research papers on power consumption

ORIGINAL RESEARCH article

Power consumption predicting and anomaly detection based on transformer and k-means.

• 1 Data Mining Laboratory, College of Mathematics and Information Technology, Hebei University, Baoding, China
• 2 The Sixty-Third Research Institute, National University of Defense Technology, Nanjing, China
• 3 School of Economics, Zhejiang University of Finance and Economics, Hangzhou, China

With the advancement of technology and science, the power system is getting more intelligent and flexible, and the way people use electric energy in their daily lives is changing. Monitoring the condition of electrical energy loads, particularly in the early detection of aberrant loads and behaviors, is critical for power grid maintenance and power theft detection. In this paper, we combine the widely used deep learning model Transformer with the clustering approach K-means to estimate power consumption over time and detect anomalies. The Transformer model is used to forecast the following hour’s power usage, and the K-means clustering method is utilized to optimize the prediction results, finally, the anomalies is detected by comparing the predicted value and the test value. On real hourly electric energy consumption data, we test the proposed model, and the results show that our method outperforms the most commonly used LSTM time series model.

1 Introduction

Modern power systems are evolving in a more sustainable path. The load demand for domestic electrical energy is gradually increasing as the number of household appliances and electric cars increases. Statistics show that residences and commercial buildings account for three-fifths of global electricity use ( Desai, 2017 ). The power system has grown in complexity and intelligence, and more modern information transmission technology has been implemented, making grid processing more convenient and secure ( Bayindir et al., 2016 ). Moreover, electric energy consumption in everyday living is also difficult and variable. Electric energy usage, for example, may vary significantly depending on the season, and consumption on working days and working days will fluctuate. At the same time, there will be anomalies in the electrical load, such as forgetting to turn off electrical appliances, failure of electrical appliances and even the theft of electricity, and so on, resulting in a much larger electrical demand than typical. As a result, detecting unusual consumption data is critical.

Abnormal detection can enhance abnormal electric energy consumption to achieve energy savings, remind users to discover malfunctioning electrical appliances or modify bad electricity usage patterns, lower users’ energy consumption expenses, and promote electricity consumption safety awareness. The most crucial factor is that you can locate the source of the power theft ( McLaughlin et al., 2009 ). According to the survey, power theft accounts for about half of the energy lost in some developing countries ( Antmann , 2009 ), and anomaly detection technologies can successfully combat this scenario.

Anomaly detection, as the name suggests, is the method of recognizing data that differs from the usual. Anomalies in data are situations that do not follow the specified usual behavior pattern ( Chandola et al., 2009 ). Anomalies are classified into three types: point anomalies, aggregate anomalies, and context anomalies. A point anomaly occurs when one point in the data is excessively high or too low in comparison to other points. The anomalous phenomena of a group of data compared to the full data set is referred to as a collection anomaly, and it only happens in the data set with the correlation between the data. Contextual abnormality refers to the abnormality when the data is combined with the context in the data set ( Chandola et al., 2009 ). In this paper, abnormal power consumption means that if the difference between the power consumption predicted by the model and the real power consumption in a certain hour is greater than the threshold we set through the experiment, the current hour power consumption is considered abnormal, so the main task of this paper is to detect point anomalies.

Because the characteristics of variables are various, traditional models primarily focus on univariate prediction and anomaly detection ( Hu et al., 2018 ). Univariate models are typically utilized in cases where there are too many other features or when vectorization is difficult, such as stock prediction ( Hsieh et al., 2011 ). Various industries, such as speech recognition ( Graves et al., 2013 ) and NLP (Natural Language Processing) ( Nadkarni et al., 2011 ), have adopted deep learning technology and achieved remarkable success as a result of the rapid development of the field of deep learning. Time series analysis ( Kuremoto et al., 2014 ), of course, has also a significant advancement. Traditional statistical methods such as ARIMA (Auto-Regressive Integrated Moving Average) ( Yuan et al., 2016 ) and SARIMA (Seasonal ARIMA) ( Ahn et al., 2015 ) were defeated by the proposed LSTM (Long and Short-Term Memory network) ( Hochreiter and Schmidhuber , 1997 ). For energy consumption prediction and anomaly detection, a lot of work on LSTM has been done.

However, with the introduction of Google’s Transformer model ( Vaswani et al., 2017 ), this model was first successfully used to the field of machine translation, and then it spread to other significant fields such as speech recognition ( Wang et al., 2020 ), and so on. Since machine translation technology involves the processing of time series, we seem to be able to use the Transformer model for time series forecasting tasks. Transformer uses self-attention and multi-head self-attention for semantic extraction. When it comes to the long-distance dependence of features in time series, self-attention can naturally solve this problem, because there are connections between all features of time series when integrating features, and the relative position information between the input time series features is preserved through sinusoidal position encoding. It is not like RNN (Recurrent Neural Network) that needs to be gradually passed to the back through hidden layer node sequences, nor is it like CNN (Convolutional Neural Networks) that needs to be captured long distance features by increasing the network depth, Transformer has some advantages in processing time series features.

As a result, in this paper, we propose a new power consumption prediction and anomaly detection model that combines deep learning and clustering methods. The following are the contributions:

1) For time series prediction of power consumption, we merged the current popular Transformer deep learning model with K-means clustering. We reasoned that the historical time data contributes differentially to the expected value due to the regular behavior of household users. The K-means method can be used to locate data clusters that contribute more to the projected value, allowing the Transformer model prediction value to be optimized further.

2) In the experiment, we employed multi-dimensional data. The data dimension also incorporates auxiliary information data such as voltage, current, and the power consumption of various household appliances, in addition to the fundamental power consumption.

3) We compared the proposed method to the LSTM and only Transformer model’s prediction performance. Experiments have revealed that the proposed combination method’s error between predicted and true values is lower than those of LSTM and single Transformer.

4) To demonstrate the proposed method’s superior performance in anomaly detection, we manually added anomalous data into the test data and treated it as a true anomaly.

The following is how we organize this paper. We introduce relevant research on power consumption and anomaly detection in Section 2. The data set used in the experiment, as well as the data set’s preparation approach, are shown in Section 3. We describe our model’s implementation approach and procedure in detail in Section 4. We compare the performance of model prediction and anomaly detection with different models in Section 5. This paper was summarized in the last section.

2 Related Works

Researchers have done a lot of research since power consumption prediction and anomaly detection are so crucial in the power energy system. Box et al. (2015) developed time series forecasting approaches like as Auto-Regressive (AR), ARIMA (Auto-Regressive Integrated Moving Average), and SARIMA (Seasonal ARIMA) in the economic sphere, and they had good results. To predict the value at a specific moment in the time series, the AR model primarily uses the weighting of all values preceding that time. ARIMA primarily employs the point before to the time to add a random vector in order to forecast the value at that time. SARIMA is mostly used for time series data with obvious seasonal differences. Ouyang et al. (2019b) improved the performance of wind power ramp prediction by combining the advantages of AR models and Markov chain. To detect anomalies, Yan et al. (2014) integrated the AR approach with SVM (Support Vector Machine) Ma and Guo (2014) . ARIMA was used by Ediger and Akar (2007) to forecast fuel energy use in Turkey. The time series of ARIMA power consumption was utilized by Alberg and Last (2018) to estimate future power consumption, and Krishna et al. (2015) employed ARIMA for half-hour granular power consumption data. SARIMA was applied by Ahn et al. (2015) for long-term and mid-term load forecasting. The unsupervised K-means approach Münz et al. (2007) groups the data to identify anomalies that are outside of the cluster. Simultaneously, the autoencoder model has been a huge success. The data is analyzed using unsupervised methods. The difference between input and output is utilized to detect whether the data is aberrant, from compression and abstraction to recovery and rebuilding. For anomaly detection, Al-Abassi et al. (2020) presented unsupervised stacked autoencoders for smart cyber-physical grids. Deb et al. (2015) developed an artificial neural network for predicting building energy usage in Singapore, and it was shown to be accurate.

The advancement of deep learning has improved the accuracy and performance of large data processing and prediction. Deep learning was used extensively in wind speed prediction ( Khodayar et al., 2017 ), stock prediction ( Rather et al., 2015 ), automated Vehicles ( Shen et al., 2020 ) and other researches, and power grid technology has also incorporated the nerve of deep learning. The network is used to forecast and detect how much energy the user consumes. Ouyang et al. (2019a) proposed the use of Deep Belief Network (DBN) to predict hourly power load. According to Shi et al. (2017) , predicting the electricity usage of a single customer in Ireland is the same as using a deep recursive network. For time series, LSTM can forecast and detect anomalies ( Malhotra et al., 2016 ; Siami-Namini et al., 2018 ). Wang et al. (2019) proposed combining seasonal features with LSTM for power load forecasting and anomaly detection. However, because ARIMA requires time series data to be stationary and it can only capture linear relationships, but not non-linear relationships. For the LSTM model, its output at the current time requires not only the input at the current time, but also the output at the previous time. This makes the LSTM model unable to parallelize operations, resulting in too long training time when processing time series features. On the other hand, the Transformer model has had a lot of success in the field of speech recognition and natural language processing since it was introduced. As a result, we propose in this paper that we utilize the transformer model to estimate electric energy load, then apply the k-means approach to further improve the prediction results, and then compare the prediction results to the test data to look for anomalies.

For the experiment, we used hour-level electricity load data from a French family for 1,440 days (2006-12-17 to 2010-11-25). We selected 3 h of data for display, as shown in Table 1 , except for “global_active_power” represents the total active power consumed by the household, and other data includes “global_reactive_power” representing the total reactive power consumed by the household, “voltage” representing the average voltage per hour, “global_intensity” representing the average current intensity, “sub_metering_1” representing the active electrical energy of the kitchen, “sub_metering_2” representing the active energy of the laundry, “sub_metering_3” representing the active energy of the climate control system, “sub_metering_4” representing other active energy. The hourly power load change diagram for 3 days which are all weekdays is shown in Figure 1 . It can be seen that power consumption has increased significantly in the morning, noon and evening, and electricity consumption conforms to the law of electricity consumption in French households during workdays. To make the data more stable, we apply the Min-Max Normalization procedure. This will make the model’s training easier and its convergence faster. We designed the model supervision task to estimate the following hour’s electric energy usage based on multivariate data collected every 23 h, and we implemented it by using a 23-hour sliding window.

TABLE 1 . Sample display of the data set used in this paper.

FIGURE 1 . Hourly household electric energy consumption change diagram for 3 consecutive days.

4 Methodology

We partitioned the data into 24-hour groups using a sliding window, then trained k-means clustering for the first 23 h of each group of real test data into k clusters, while also used the 23-hour real load data training Transformer model predicts the next hour’s load data, then through the trained K-means to get the appropriate centroid as the final prediction result. Figure 2 depicts the framework of our model.

FIGURE 2 . The main framework of our model.

4.1 Transformer Model

Initial and foremost, Positional Encoding is the first phase in the Transformer model utilized in this essay. Because Transformer does not have a cyclic structure like LSTM, it presents a new positional encoding strategy to capture the input time series information, as indicated in Eqs 1, 2 . The basic idea is to add sine and cosine functions of various frequencies as position codes to the normalized input sequence, allowing the Transformer model’s multi-head attention mechanism to fully capture time series data with more dimensions.

where pos is the vector position of each time. For example, in the time series data in this paper, the pos of the first hour of each group is 0, and the pos of the second hour is 1, 2 i and 2 i + 1 respectively represent the even position and the odd position. d model represents the length of the feature vector per hour. Next, we use X = [ x 1 , x 2 , …, x n ] to represent the input sequence combined with position encoding, and pass the multi-head self-attention of the Transformer model:

In the above formula, Q = q 1 w q 1 , K = k 1 w k 1 , V = v 1 w v 1 , q 1 = k 1 = v 1 = X . In the Transformer model, the Attention module first undergoes a linear transformation of Q (Query), K (Key), and V (Value). Each time Q, K, and V perform the linear transformation, the parameter W is different, and then input to Scaling dot product attention, the formula is as (5), note that it is necessary to do i times here, in fact, it is the so-called multi-head, and each time counts as one head. Then concatenate the attention results of the i times of scaling dot product, and then perform a linear transformation to obtain the value as the result of the multi-head attention. The advantage of this is that it allows the model to learn relevant information in different representation subspaces. The calculation of the Attention module uses Scaled Dot-product:

where d k is the last dimension of the shape of Q , K , V , divided by d k to prevent d k from being too large and the softmax function’s gradient becoming too tiny when QK T is too large. The residual connection structure is then used to narrow the network’s attention to solely the differences. In multi-layer network structures, it is frequently used:

where LayerNorm is Layer Normalization, which normalizes each neuron and adjusts the mean and variance of the input data to be the same, which will speed up the convergence. Then input L into the FeedForward layer, which is composed of two fully connected layers, the first layer uses the Relu activation function, and the second layer does not use the activation function:

Similarly, use residual connection and Layer Normalization again to get the output S E :

Our experiment uses a 2-layer Transformer multi-head attention module, which means that the output S E needs to be re-input to the structure output S E 2 described above. Finally, the output will be decoded and dimensionality reduction operations through the fully connected layer:

4.2 K-Means Clustering Method

Clustering is the division of a sample set into several disjoint subsets (sample clusters), which is a typical unsupervised machine learning algorithm. When using clustering to classify samples, Euclidean Distance is used as the measurement criterion of sample similarity. The higher the similarity, the smaller the Euclidean Distance of the sample. K-means clustering is a well-known algorithm among clustering algorithms. It needs to determine the number of clusters k first when clustering, and k is given by the user. Each cluster passes through its centroid (the mean value of all elements in the cluster). The workflow of k-means is also very simple. First, randomly select k initial points as the initial centroids of each cluster, and then assign each point in the data set to the cluster closest to it. The distance calculation uses the Euclidean Distance mentioned above. The algorithm of k-means is shown in Algorithm 1 :

Algorithm 1 K -means algorithm.

4.3 Model Development

First, we divide the data set into a training set and a test set. Since the data set contains a total of 1,440 days of hourly data, we choose 1,240 days of data as the training set and 200 days of data as the test set. For the Transformer model, we choose two consecutive layers of multi-head self-attention modules, and each multi-head attention is set to 4 attention heads. The input data shape is 23 time steps and 8 features. For K-means clustering, we experimented with k = 2, 5, 8, 10, 11, and 15 respectively, and finally selected the cluster with k = 10. We choose the mean-square error of the predicted data and the original data, that is, MSE (Mean-Square Error) as the loss function, and Adam as the optimizer of the model. And set the epoch of the training model to 300, and the batch size to 200.

4.4 Model Prediction and Evaluation

We believed that the nearest neighbor of the real training data has the most impact on the forecast value, thus in the first 23 h of each group, we trained k-means clustering, partitioned the data into k categories, and provided the load K predicted by the Transformer model for the next hour as the final prediction output, then found the centroid in the trained K-means cluster. We analyzed the model’s prediction ability by fitting the predicted value to the test value and calculating the RMSE (Root Mean Squard Error) of the forecasted value and the test value to measure the prediction’s accuracy, and we compared it to the commonly used LSTM model.

4.5 Anomaly Detection

Because the model calculates the anticipated value based on a huge quantity of historical data, the forecasted values will generally follow the data’s trend. If the test value differs significantly from the projected value, it indicates that the test value has deviated too far from the data trend and may be abnormal. The score between the predicted value and the test value obtained after the Transformer model and k-means clustering is calculated in this research. The formula can be found in Eq. 10 . In order to better analyze the difference, we normalized the score, the formula is as Eq. 11 . The value of score collected from several experiments was used to determine a threshold. When the score between the predicted value and the test value exceeds the threshold, the test value is considered abnormal. The experiment can also evaluate whether the user is prone to having electricity theft by setting a time series window and a threshold for the number of anomalies. If the number of abnormalities in the time series data in a window is greater than the threshold, it means that the time series data are abnormal, and the user may have the suspicion of steal electricity. To better compare the accuracy of anomaly detection, we manually insert abnormal data points in the test data and compare our model with K-means and LSTM.

5 Experiment Results

5.1 consumption prediction.

We opted to compare the model against the most popular LSTM model for time series data prediction in order to test its performance. The LSTM is a variant of the recurrent neural network RNN. It is a unique RNN that incorporates three different types of gating to address the problem of gradient disappearance and explosion during lengthy sequence training. Simply put, LSTM outperforms standard RNNs in longer sequences, making it ideal for time series forecasting jobs.

Figure 3 shows how the training and test loss of the Transform model used in this paper changes at 300 epoch. It can be observed that the model converges quickly, and the figure shows that there is no overfitting in the model. All of this is achievable because of the Transformer model’s benefits in time series processing.

FIGURE 3 . Train and test loss over the 300 epochs of our model.

Figure 4 depicts the test set’s real-time power consumption data over 3 days, as well as a comparison of our model’s and the LSTM model’s prediction results on the test set. The blue line represents the actual data, the red line represents our model’s predicted value, the green line represents the LSTM model’s predicted value, and the purple line indicates the lone Transformer’s predicted value. Our model’s forecast data is more in accordance with the real test data, as can be shown. The RMSE of the model, on the other hand, are used to assess the model’s fit. Our model has an RMSE of 0.73, the Transformer has an RMSE of 0.77, and the LSTM has an RMSE of 0.86. In terms of prediction accuracy, our model outperforms LSTM by 15%, while Transformer outperforms LSTM by just 10%. After our analysis, because the dimensionality of the feature vector of each time series in our time series data is too small, which leads to the failure of the full performance of the Transformer model.

FIGURE 4 . Comparison of our model and LSTM predicted value fitting real data.

5.2 Anomaly Detection

After a lot of testing and tweaking, we ultimately settled on 0.45 as the threshold. This means that any point with a score higher than 0.45 will be considered anomalous. The change in score data over 2,500 h is presented in Figure 5 , with the red dashed line representing the threshold and the purple point representing the abnormal point. The data scores are primarily focused between 0 and 0.3, and there are relatively few aberrant spots, as can be shown. In practical applications, we can adjust the threshold size based on the scene being used, and lower or increase the threshold size based on the strictness of anomaly detection, a lower threshold is more stringent, allowing for the detection of more anomalies, on the other hand, a higher threshold is more tolerant, allowing for the detection of fewer anomalies.

FIGURE 5 . Scores and anomalies exceeding the threshold for 2,500 h.

We utilized the strategy of randomly inserting abnormal points in the test data to better compare and assess the model’s anomaly detection capabilities because this experimental data set does not mark aberrant time points. In the 200 days (4,800 h) of the test set, we randomly selected a value every day and double it, and assume it is an outlier, so there are 200 outliers in the 4,800 data in the test set. For comparison, we separately used the clustering method K-means and the most popular depth method LSTM to detect abnormal points. Using the K-means approach, we discovered a total of 68 abnormal points, of which only 56 were the abnormal points we manually inserted into the data set. Using the LSTM model, we retrieved a total of 162 anomalies, 120 of which were the anomaly points we manually inserted into the data set. Our combined Transformer and K-means model found 165 anomalies, 132 of which were the abnormal points we manually added to the data set. The accuracy, precision, recall, and F1 of the three models are shown in Table 2 . The predicted RMSE of LSTM and our model are also shown in the table.

TABLE 2 . Comparison with some methods.

6 Conclusion

The prediction of electric energy consumption and the identification of anomalies are critical in the functioning of the power grid, and the processing of multi-variable time series is a difficult challenge. We present a model that combines Transformer and K-means approaches in this article. Every 23 h of training data is separated into k clusters using K-means clustering. At the same time, this training data are used to train the Transformer model to predict the following hour’s power usage, with the predicted value being placed into the trained K-means cluster and the cluster’s centroid serving as the final predicted value. Finally, look for anomalies by comparing the anticipated value to the actual test results. The experimental results prove that the model achieves prediction accuracy with less error and high anomaly detection performance. In the future, we’ll strive to improve prediction and anomaly detection accuracy, as well as study the differences between power consumption prediction and anomaly detection in different seasons, environments, and other scenarios, and other issues that need to be addressed.

Data Availability Statement

The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found below: https://archive.ics.uci.edu/ml/datasets/individual+household+electric+power+consumption .

Author Contributions

JZ: conceptualization, methodology, data preprocessing, and writing-original draft preparation. HZ: visualization, investigation. SD: experimental training and testing. XZ: supervision and reviewing.

This work was supported in part by the National Natural Science Foundation of China under Grants 71901215, 71901191, the National University of Defense Technology Research Project ZK20-46, and the Outstanding Youth Talents Program of National University of Defense Technology.

Conflict of Interest

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Publisher’s Note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Ahn, B.-H., Choi, H.-R., and Lee, H.-C. (2015). Regional Long-Term/mid-Term Load Forecasting Using Sarima in south korea. J. Korea Academia-Industrial cooperation Soc. 16, 8576–8584. doi:10.5762/kais.2015.16.12.8576

CrossRef Full Text | Google Scholar

Al-Abassi, A., Sakhnini, J., and Karimipour, H. (2020). “Unsupervised Stacked Autoencoders for Anomaly Detection on Smart Cyber-Physical Grids,” in 2020 IEEE International Conference on Systems, Man, and Cybernetics (SMC) ( IEEE ), 3123–3129. doi:10.1109/smc42975.2020.9283064

Alberg, D., and Last, M. (2018). Short-term Load Forecasting in Smart Meters with Sliding Window-Based Arima Algorithms. Vietnam J. Comput. Sci. 5, 241–249. doi:10.1007/s40595-018-0119-7

Antmann, P. (2009). Reducing Technical and Non-technical Losses in the Power Sector .

Google Scholar

Bayindir, R., Colak, I., Fulli, G., and Demirtas, K. (2016). Smart Grid Technologies and Applications. Renew. Sustain. Energ. Rev. 66, 499–516. doi:10.1016/j.rser.2016.08.002

Box, G., Jenkins, G. M., Reinsel, G. C., and Ljung, G. M. (20152015). Time Series Analysis: Forecasting and Control .

Chandola, V., Banerjee, A., and Kumar, V. (2009). Anomaly Detection. ACM Comput. Surv. 41, 1–58. doi:10.1145/1541880.1541882

Deb, C., Eang, L. S., Yang, J., and Santamouris, M. (2015). Forecasting Energy Consumption of Institutional Buildings in singapore. Proced. Eng. 121, 1734–1740. doi:10.1016/j.proeng.2015.09.144

Desai, B. H. (2017). 14. United Nations Environment Program (Unep). Yearb. Int. Environ. L. 28, 498–505. doi:10.1093/yiel/yvy072

Ediger, V. Ş., and Akar, S. (2007). Arima Forecasting of Primary Energy Demand by Fuel in turkey. Energy policy 35, 1701–1708. doi:10.1016/j.enpol.2006.05.009

Graves, A., Mohamed, A.-r., and Hinton, G. (20132013). IEEE International Conference on Acoustics, Speech and Signal Processing . IEEE , 6645–6649. doi:10.1109/icassp.2013.6638947Speech Recognition with Deep Recurrent Neural Networks

Hochreiter, S., and Schmidhuber, J. (1997). Long Short-Term Memory. Neural Comput. 9, 1735–1780. doi:10.1162/neco.1997.9.8.1735

PubMed Abstract | CrossRef Full Text | Google Scholar

Hsieh, T.-J., Hsiao, H.-F., and Yeh, W.-C. (2011). Forecasting Stock Markets Using Wavelet Transforms and Recurrent Neural Networks: An Integrated System Based on Artificial Bee colony Algorithm. Appl. soft Comput. 11, 2510–2525. doi:10.1016/j.asoc.2010.09.007

Hu, M., Ji, Z., Yan, K., Guo, Y., Feng, X., Gong, J., et al. (2018). Detecting Anomalies in Time Series Data via a Meta-Feature Based Approach. Ieee Access 6, 27760–27776. doi:10.1109/access.2018.2840086

Khodayar, M., Kaynak, O., and Khodayar, M. E. (2017). Rough Deep Neural Architecture for Short-Term Wind Speed Forecasting. IEEE Trans. Ind. Inf. 13, 2770–2779. doi:10.1109/tii.2017.2730846

Krishna, V. B., Iyer, R. K., and Sanders, W. H. (2015). “Arima-based Modeling and Validation of Consumption Readings in Power Grids,” in International Conference on Critical Information Infrastructures Security ( Springer ), 199–210.

Kuremoto, T., Kimura, S., Kobayashi, K., and Obayashi, M. (2014). Time Series Forecasting Using a Deep Belief Network with Restricted Boltzmann Machines. Neurocomputing 137, 47–56. doi:10.1016/j.neucom.2013.03.047

Ma, Y., and Guo, G. (2014). Support Vector Machines Applications , Vol. 649. Springer .

Malhotra, P., Ramakrishnan, A., Anand, G., Vig, L., Agarwal, P., and Shroff, G. (2016). Lstm-based Encoder-Decoder for Multi-Sensor Anomaly Detection. arXiv preprint arXiv:1607.00148.

McLaughlin, S., Podkuiko, D., and McDaniel, P. (2009). “Energy Theft in the Advanced Metering Infrastructure,” in International Workshop on Critical Information Infrastructures Security ( Springer ), 176–187.

Münz, G., Li, S., and Carle, G. (2007). “Traffic Anomaly Detection Using K-Means Clustering,” in GI/ITG Workshop MMBnet , 13–14.

Nadkarni, P. M., Ohno-Machado, L., and Chapman, W. W. (2011). Natural Language Processing: an Introduction. J. Am. Med. Inform. Assoc. 18, 544–551. doi:10.1136/amiajnl-2011-000464

Ouyang, T., He, Y., Li, H., Sun, Z., and Baek, S. (2019a). Modeling and Forecasting Short-Term Power Load with Copula Model and Deep Belief Network. IEEE Trans. Emerg. Top. Comput. Intell. 3, 127–136. doi:10.1109/tetci.2018.2880511

Ouyang, T., Zha, X., Qin, L., He, Y., and Tang, Z. (2019b). Prediction of Wind Power Ramp Events Based on Residual Correction. Renew. Energ. 136, 781–792. doi:10.1016/j.renene.2019.01.049

Rather, A. M., Agarwal, A., and Sastry, V. N. (2015). Recurrent Neural Network and a Hybrid Model for Prediction of Stock Returns. Expert Syst. Appl. 42, 3234–3241. doi:10.1016/j.eswa.2014.12.003

Shen, X., Zhang, X., Ouyang, T., Li, Y., and Raksincharoensak, P. (2020). Cooperative Comfortable-Driving at Signalized Intersections for Connected and Automated Vehicles. IEEE Robot. Autom. Lett. 5, 6247–6254. doi:10.1109/lra.2020.3014010

Shi, H., Xu, M., and Li, R. (2017). Deep Learning for Household Load Forecasting a Novel Pooling Deep Rnn. IEEE Trans. Smart Grid 9, 5271–5280.

Siami-Namini, S., Tavakoli, N., and Namin, A. S. (2018). “A Comparison of Arima and Lstm in Forecasting Time Series,” in 2018 17th IEEE International Conference on Machine Learning and Applications (ICMLA) ( IEEE ), 1394–1401. doi:10.1109/icmla.2018.00227

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., et al. (2017). Attention Is All You Need. Advances in Neural Information Processing Systems , 5998–6008.

Wang, X., Zhao, T., Liu, H., and He, R. (20192019). “Power Consumption Predicting and Anomaly Detection Based on Long Short-Term Memory Neural Network,” in IEEE 4th international conference on cloud computing and big data analysis (ICCCBDA) ( IEEE ), 487–491. doi:10.1109/icccbda.2019.8725704

Wang, Y., Mohamed, A., Le, D., Liu, C., Xiao, A., Mahadeokar, J., et al. (2020). “Transformer-based Acoustic Modeling for Hybrid Speech Recognition,” in ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) ( IEEE ), 6874–6878. doi:10.1109/icassp40776.2020.9054345

Yan, K., Shen, W., Mulumba, T., and Afshari, A. (2014). Arx Model Based Fault Detection and Diagnosis for Chillers Using Support Vector Machines. Energy and Buildings 81, 287–295. doi:10.1016/j.enbuild.2014.05.049

Yuan, C., Liu, S., and Fang, Z. (2016). Comparison of China's Primary Energy Consumption Forecasting by Using ARIMA (The Autoregressive Integrated Moving Average) Model and GM(1,1) Model. Energy 100, 384–390. doi:10.1016/j.energy.2016.02.001

Keywords: power consumption prediction, anomaly detection, transformer, K-means, LSTM

Citation: Zhang J, Zhang H, Ding S and Zhang X (2021) Power Consumption Predicting and Anomaly Detection Based on Transformer and K-Means. Front. Energy Res. 9:779587. doi: 10.3389/fenrg.2021.779587

Received: 19 September 2021; Accepted: 06 October 2021; Published: 22 October 2021.

Reviewed by:

Copyright © 2021 Zhang, Zhang, Ding and Zhang. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

*Correspondence: Song Ding, [email protected]

Disclaimer: All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

• View all journals
• Explore content
• About the journal
• Publish with us
• Sign up for alerts
• Open access
• Published: 02 January 2024

Research on the interaction between energy consumption and power battery life during electric vehicle acceleration

• Qin Liu 1 ,
• Zhongbo Zhang 1 &
• Jingjing Zhang 1  

Scientific Reports volume  14 , Article number:  157 ( 2024 ) Cite this article

1298 Accesses

2 Citations

Metrics details

• Electrical and electronic engineering
• Energy science and technology

Most studies on the acceleration process of electric vehicle focus on reducing energy consumption, but do not consider the impact of the power battery discharge current and its change rate on the battery life. Therefore, this paper studied the interaction between electric vehicle energy consumption and power battery capacity attenuation during acceleration. First, a power battery life model for electric vehicle under driving conditions is established, and the percentage of battery capacity loss per kilometer is used to measure the capacity loss under different acceleration conditions. Then, the relationship between the percentage of battery capacity loss per kilometer and velocity and acceleration is explored, and the capacity attenuation mechanism of power battery under different acceleration processes is analyzed. Finally, the energy consumption and battery capacity attenuation is studied when the electric vehicle accelerated with multiple accelerations curves, and the interaction of the first acceleration and acceleration time on the electric vehicle energy consumption and the power battery capacity attenuation characteristics is discussed. The research results indicate that when the electric vehicle accelerates with different multiple accelerations curves, the change of energy consumption per kilometer and percentage of battery capacity loss per kilometer with acceleration and acceleration time is different, and the change of the two is basically opposite.

Similar content being viewed by others

Trade-offs between automation and light vehicle electrification

Ecological driving on multiphase trajectories and multiobjective optimization for autonomous electric vehicle platoon

Analysis of the influence of passenger load on bus energy consumption a vehicle-engine combined model-based simulation framework

Introduction.

The existing research on the electric vehicle (EV) energy consumption mainly focuses on the matching of powertrain parameters, the development of efficient control strategies for motor drive systems, and the optimization control strategies for energy management, and so on. For example, in the aspect of parameter matching, it mainly includes improving the working efficiency of each component of the powertrain system and reducing EV energy consumption, the parameters of the electric motor, battery and its transmission system are matched reasonably 1 , 2 , 3 , 4 . In the aspect of motor efficiency optimization control strategy, so as to reduce EV energy consumption and improve driving range, a series of motor efficiency optimization control strategies have been developed to achieve efficient operation of motors in high power range 5 , 6 , 7 . In the aspect of energy management optimization control strategy, fuzzy logic, dynamic planning and other control strategies are used to reduce electric vehicle energy consumption rate, thereby reducing EV energy consumption and increasing the driving range 8 , 9 , 10 , 11 .

Moreover, the research in references 12 , 13 found that the magnitude of acceleration has a significant impact on EV energy consumption. However, there are few existing studies in EVs’ acceleration. For instance, in reference 14 , the difference of EV energy consumption under a single acceleration value and multiple accelerations values is studied. The research indicates that, compared to a single acceleration, when the EV accelerates with multiple accelerations values, its energy consumption is less. In reference 15 , the image of driving mode on the EV energy consumption is studied, and it is found that the energy consumption is lower when the EV is driven in a driving mode with a lower average velocity. Besides, the impact of different environmental and control parameters on energy consumption and driving range characteristics is studied in 16 ; the research results indicate that the main factors affecting energy consumption and driving distance are average vehicle velocity, driving time, etc. Therefore, drivers can obtain more energy-efficient routes by estimating the driving distance under different driving conditions. And the relationship between EV energy consumption and acceleration time is discussed in 17 , and the results show that when the acceleration time is extended within an appropriate range, the energy consumption can be effectively reduced, and the lower the speed, the greater the energy saving potential.

And in terms of the analysis of power battery life, the main focus is on the prediction of power battery cycle life model. There are three prediction methods for power battery life: model-based prediction, data-driven prediction and fusion technology prediction 18 , 19 , 20 . Such as, an empirical model is used in reference 21 to model the global and local degradation of lithium-ion battery aging process. In 22 , a physics model-based method is designed to predict battery capacity and remaining useful life, using a semi-empirical model, degradation parameters are estimated from voltage and capacity measurements to predict capacity decay trends. And in 23 , a new customary model is proposed to predict the battery degradation trend, the Bayesian inference technology is used to estimate the model parameters, particle filter step by step estimation and Markov chain Monte Carlo algorithm batch estimation, and then the estimated parameters are used to generate transient events and predict future capacity degradation. In addition, in reference 24 , a new degradation prediction model that integrates the effects of temperature and charge discharge cycles on battery degradation is established, and the PSO algorithm is applied to optimize the model parameters. And in reference 25 , the data-driven method based on discrete wavelet transform is proposed to prediction the remaining useful life of lithium-ion battery. Besides, Ahwiadi et al. 26 used enhanced particle filter to improve the modeling accuracy of battery health monitoring and remaining useful life prediction. And in order to improve prediction accuracy, Duan et al. 27 used the extended Kalman filter as sampling density function to optimize particle filter algorithm. The fusion method can be divided into two types: model and data-driven fusion method and different data-driven fusion method. For instance, Wei et al. 28 used the fusion method of optimal correlation vector and improved degradation model with Hausdorff distance to predict the remaining useful life of lithium-ion battery. Wang et al. 29 used the hybrid model of support vector regression and differential evolution to predict the remaining useful life of lithium ion battery. Zhu et al. 30 used battery degradation data and Bayesian updating algorithm to establish the based-performance degradation Wiener process model. And in reference 31 , a hybrid method based on bidirectional long short-term memory model with attention mechanism and support vector regression model is proposed for online battery cycle life prediction.

On the basis of previous research, on the one hand, many studies are put into effect in reducing EV energy consumption by different control strategies, and the driving mode and influencing factors have also been analyzed. On the other hand, most research on power batteries mainly focuses on its remaining useful life. However, few studies have combined acceleration and acceleration time with EV energy consumption and battery life. Therefore, in this paper, the interaction mechanism between energy consumption and power battery life during the acceleration process of EVs is studied. The impact of vehicle velocity and acceleration on energy consumption and battery life is analyzed, considering the characteristic of the discharge rate of power batteries used in EVs constantly changing with the driving conditions, the capacity attenuation model of power battery under driving conditions is established, and the effect of discharge current and its change on battery capacity attenuation is discussed. In addition, when the EV accelerates with convex acceleration curves with multiple accelerations values, the interaction relationship between energy consumption and power battery capacity attenuation is also studied, and the variation of energy consumption and battery life with acceleration and acceleration time is analyzed. The research in this paper can provide guidance for the optimization of the EV acceleration mode, so as to further reduce the energy consumption and thus prolong the battery life.

Mathematical model of EV energy consumption and battery life

The model of ev energy consumption during accelerating process.

According to the principle of longitudinal dynamics of automobile 32 , when the EV accelerates on good road, the output power of its power battery \(P_{{\text{b}}} (t)\) \(\left( {\text{W}} \right)\) at time t can be obtained as:

where the \(F_{{\text{t}}} (t)\) (N) is traction force; \(\eta_{T}\) is the efficiency of the transmission system; \(\eta_{{\text{V}}}\) is the efficiency of the inverter; \(\eta_{{\text{m}}}\) is the efficiency of the electric motor.

Therefore, when an EV accelerates according to different acceleration conditions, the energy consumption per kilometer \(E_{{{\text{b}} - {\text{j}}}}^{l} \left( t \right)\) in the \(l\) acceleration range \(\left[ {t_{l} \sim t_{l + 1} } \right]\) is:

By bringing the traction force into the above formula, the EV energy consumption per kilometer of the \(l\) acceleration range \(\left[ {t_{l} \sim t_{l + 1} } \right]\) can be obtained as follows:

The above equation can also be expressed as:

In Eq. ( 4 ), \(k_{{\text{f}}}\) is defined as the impact factor of road resistance, \(k_{{\text{f}}} = {{\left( {mgf + mgi} \right)} \mathord{\left/ {\vphantom {{\left( {mgf + mgi} \right)} {3.6\eta_{{\text{T}}} \eta_{{\text{V}}} \eta_{{\text{m}}} }}} \right. \kern-0pt} {3.6\eta_{{\text{T}}} \eta_{{\text{V}}} \eta_{{\text{m}}} }}\) ; \(k_{{{\text{C}}_{{\text{D}}} }}\) is defined as the impact factor of air resistance, \(k_{{{\text{C}}_{{\text{D}}} }} = {{C_{D} A} \mathord{\left/ {\vphantom {{C_{D} A} {76.14 \cdot \eta_{{\text{T}}} \eta_{{\text{V}}} \eta_{{\text{m}}} }}} \right. \kern-0pt} {76.14 \cdot \eta_{{\text{T}}} \eta_{{\text{V}}} \eta_{{\text{m}}} }}\) ; \(k_{{\updelta }}\) is defined as the impact factor of acceleration resistance, \(k_{{\updelta }} = {{\delta m} \mathord{\left/ {\vphantom {{\delta m} {3.6\eta_{{\text{T}}} \eta_{{\text{V}}} \eta_{{\text{m}}} }}} \right. \kern-0pt} {3.6\eta_{{\text{T}}} \eta_{{\text{V}}} \eta_{{\text{m}}} }}\) .

Then, the EV energy consumption per kilometer \(E_{{\text{b - j}}} \left[ {u\left( t \right),a\left( t \right)} \right]\) during the whole acceleration process can be expressed as:

From Eqs. ( 4 ) and ( 5 ), it can be concluded that during the acceleration process of an EV, when the vehicle parameters and road conditions are constant, its \(k_{{\text{f}}}\) , \(k_{{{\text{C}}_{{\text{D}}} }}\) and \(k_{{\updelta }}\) can be regarded as fixed values. The \(E_{{{\text{b}} - {\text{j}}}} \left[ {u\left( t \right),a\left( t \right)} \right]\) mainly depends on \(u\left( t \right)\) and \(a\left( t \right)\) , i.e. \(E_{{{\text{b}} - {\text{j}}}} \left[ {u\left( t \right),a\left( t \right)} \right] = f\left( {u\left( t \right),a\left( t \right)} \right)\) , while the \(u\left( t \right)\) and \(a\left( t \right)\) mainly depend on different acceleration conditions. Therefore, studying different acceleration processes and implementing reasonable control is an important way to reduce the EV energy consumption.

The model of battery life during accelerating process

During the EV acceleration process, the power battery life mainly depends on the discharge rate \(n\) , temperature \(T\) , discharge depth \(DOD\) , etc. Among domestic and foreign scholars' research on the life model of lithium-ion batteries, the model established by Wang John 33 of the American HRL laboratory is the most representative. In the study of the impact of \(T\) , \(n\) , and \(DOD\) on battery capacity, the battery capacity loss rate was used to predict the battery life, and according to the experimental results in reference 33 , the battery capacity loss model of lithium-ion power batteries can be expressed as:

where \(B\) is pre-exponential factor; \(A_{h}\) ( \({\text{A}}\;{\text{h}}\) ) represents the cumulative ampere-hours, so it can be expressed as \(A_{{\text{h}}} = \int_{0}^{t} {\left| {I\left( t \right)} \right|dt}\) , \(R\) is gas constant, \(R = 8.314\;{\text{J/}}\left( {{\text{mol}}\;{\text{K}}} \right)\) .

The Eq. ( 6 ) reflects the battery capacity loss of a single battery under constant current discharge conditions. To verify the correctness of the model and its applicability, a certain lithium-ion power battery for electric vehicle is used for charge–discharge experiments. Set the termination voltage of charge and discharge to 3.65 V and 2.5 V, and discharge at 1C and 3C, respectively. The test results are shown in Fig. 1 .

Comparison of theoretical calculation and experimental results of power battery capacity loss.

As can be seen from Fig. 1 , the percentage of battery capacity loss calculated by the model is basically consistent with the experimental results, and the experimental results of the percentage of battery capacity loss under 1C and 3C discharge are evenly distributed on both sides of the theoretical value. The model of formula ( 6 ) is only applicable to the life prediction analysis of a single battery under constant discharge rate. However, the working conditions of EVs are variable in the actual operation process, and the discharge current of its power battery varies greatly, and the corresponding discharge rate is not a constant value and fluctuates greatly. Therefore, it is necessary to revise the capacity loss model under constant discharge conditions, and refer to the method in reference 34 to predict the battery life under actual driving conditions. Based on equal life loss, assuming that the battery capacity loss rate under 1C discharge rate is \(Q_{{{\text{loss1}}}}\) , and the battery capacity loss rate at nC discharge rate is \(Q_{{{\text{loss}}n}}\) , the battery capacity loss under nC discharge condition can be deduced from \(Q_{{{\text{loss1}}}} = Q_{{{\text{loss}}n}}\) as follows:

where \(B_{1} = 27790\) , \(B_{n} { = }36042 - 9419n{ + }1215n^{2} - 48n^{3}\) 。

Since the acceleration process is studied in this paper, only the discharge process of the power battery is considered, then the capacity attenuation \(Q_{{\text{loss - j}}}^{l} \left( t \right)\) of the power battery in the \(l\) acceleration range \(\left[ {t_{l} \sim t_{l + 1} } \right]\) can be expressed as:

where \(I_{1}\) is the current at 1C discharge rate of the single battery; \(n\left( t \right)\) is the discharge rate at time t , \(n\left( t \right) = {{I\left( t \right)} \mathord{\left/ {\vphantom {{I\left( t \right)} {C_{{\text{r}}} }}} \right. \kern-0pt} {C_{{\text{r}}} }}\) , \(C_{{\text{r}}}\) is the nominal capacity of a single battery, and \(I\left( t \right)\) is the discharge current of a single battery during accelerating driving. Ignoring the impact of single battery inconsistency, assuming that the current flowing through each parallel module is consistent, the current of the single battery is \(I\left( t \right) = {{I_{{{\text{pack}}}} \left( t \right)} \mathord{\left/ {\vphantom {{I_{{{\text{pack}}}} \left( t \right)} {Num}}} \right. \kern-0pt} {Num}}\) , and \(Num\) is the parallel number of the battery pack. The current of the battery pack is \(I_{{{\text{pack}}}} \left( t \right) = I_{{\text{b}}} \left( t \right)\) , and the output current of the power battery pack \(I_{{\text{b}}} \left( t \right)\) can be expressed as \(I_{{\text{b}}} \left( t \right) = {{P_{{\text{b}}} (t)} \mathord{\left/ {\vphantom {{P_{{\text{b}}} (t)} {U_{{\text{b}}} \left( t \right)}}} \right. \kern-0pt} {U_{{\text{b}}} \left( t \right)}}\) . The voltage fluctuation of the power battery pack during the EV acceleration process is also very small, which can be regarded as a constant. Therefore, the capacity attenuation of the power battery in an acceleration condition can be obtained by substituting the corresponding parameters:

The percentage of battery capacity loss per kilometer \(q_{{{\text{loss}} - {\text{j}}}}^{l} \left[ {u\left( t \right),a\left( t \right)} \right]\left( {{\text{\% /km}}} \right)\) in an acceleration condition can be expressed as:

And the \(q_{{{\text{loss}} - {\text{j}}}} \left[ {u\left( t \right),a\left( t \right)} \right]\) of the power battery of an EV during the entire acceleration process can be expressed as:

From the above analysis, it can be concluded that during the EV acceleration process, when the vehicle parameters and road conditions are constant, the \(k_{{\text{f}}}, k_{{{\text{C}}_{{\text{D}}} }}\) and \(k_{{\updelta }}\) can also be considered as fixed values. Then the \(q_{{{\text{loss}} - {\text{j}}}} \left[ {u\left( t \right),a\left( t \right)} \right]\) mainly depends on \(u\left( t \right)\) and \(a\left( t \right)\) , that is \(q_{{{\text{loss}} - {\text{j}}}} = f\left[ {u\left( t \right),a\left( t \right)} \right]\) , and \(u\left( t \right)\) and \(a\left( t \right)\) mainly depend on the acceleration curve. Therefore, studying the different acceleration processes of EVs and implementing reasonable control is also a way to extend the battery life. And Table 1 shows the main parameters of the EV and the battery.

Research on the relationship between EV energy consumption and power battery Life with multiple acceleration curves

On the basis of research in reference 35 , for a given acceleration condition which the initial velocity and the final velocity and the total acceleration time are constant, when the EV accelerates with convex multiple accelerations curves, its energy consumption per kilometer is lower than using a single acceleration. Therefore, this section analyzes the influence of the number of acceleration and the acceleration time in the multiple accelerations curves on the EV energy consumption and the battery life under different acceleration conditions. Select three acceleration conditions with zero initial velocity and final velocity of 40 km/h, 80 km/h, and 120 km/h respectively to analyze the impact of different acceleration numbers on energy consumption per kilometer and percentage of battery capacity loss per kilometer. Among them, 0–40 km/h is defined as the neighborhood condition, 0–80 km/h is defined as the urban condition, and 0–120 km/h is defined as the highway condition. Taking into account the actual acceleration conditions, the number of accelerations cannot be infinitely large, refer to the limits of the number of accelerations in 14 , in this paper, the value of acceleration number n is \(n \le 6\) . The acceleration time and the acceleration for three working conditions are set as shown in Table 2 .

According to the setting of acceleration time and acceleration under different working conditions in the table, combined with the previous research 33 , this article mainly delves into the changes of \(E_{{\text{b - j}}} \left[ {u\left( t \right),a\left( t \right)} \right]\) and \(q_{{{\text{loss}} - {\text{j}}}} \left[ {u\left( t \right),a\left( t \right)} \right]\) with acceleration and acceleration time, when the EV accelerates with three acceleration values. Figure 2 shows the relationship between vehicle velocity and acceleration time, let the acceleration and acceleration time of the first segment be \(a_{1}^{{{\text{III}}}}\) ( \({\text{m/s}}^{{2}}\) ) and \(t_{{1}}^{{{\text{III}}}}\) (s) respectively, the acceleration and acceleration time of the second segment be \(a_{{2}}^{{{\text{III}}}}\) ( \({\text{m/s}}^{{2}}\) ) and \(t_{{2}}^{{{\text{III}}}}\) (s) respectively, the acceleration and acceleration time of the third segment be \(a_{{3}}^{{{\text{III}}}}\) ( \({\text{m/s}}^{{2}}\) ) and \(t_{{3}}^{{{\text{III}}}}\) (s) respectively, and \(a_{1}^{{{\text{III}}}} > a_{{2}}^{{{\text{III}}}} > a_{{3}}^{{{\text{III}}}}\) , and the final acceleration velocity and acceleration time be \(u_{{\text{f}}}^{{{\text{III}}}}\) ( \({\text{km/h}}\) ) and \(t_{{\text{f}}}^{{{\text{III}}}}\) (s) respectively. Assuming that \(a_{1}^{{{\text{III}}}} ,t_{1}^{{{\text{III}}}}\) is known, then \(a_{{3}}^{{{\text{III}}}}\) and \(t_{{3}}^{{{\text{III}}}}\) can be expressed by \(u_{{\text{f}}}^{{{\text{III}}}} ,t_{{\text{f}}}^{{{\text{III}}}} ,a_{1}^{{{\text{III}}}} ,t_{1}^{{{\text{III}}}} ,a_{2}^{{{\text{III}}}} ,t_{2}^{{{\text{III}}}}\) . Therefore, when the EV accelerates with a three acceleration curve, the variation of its energy consumption per kilometer \(E_{{{\text{b}} - {\text{j}}}}^{{{\text{III}}}}\) \(\left( {\text{Wh/km}} \right)\) and the percentage of battery capacity loss per kilometer \(q_{{{\text{loss}}}}^{{{\text{III}}}}\) \(\left( {{\text{\% /km}}} \right)\) with \(a_{2}^{{{\text{III}}}}\) and \(t_{2}^{{{\text{III}}}}\) can be analyzed. Considering two situations in this section, the first is to increase \(a_{1}^{{{\text{III}}}}\) when \(t_{1}^{{{\text{III}}}}\) is constant, and analyze the variation of the \(E_{{\text{b - j}}}^{{{\text{III}}}}\) and the \(q_{{{\text{loss}}}}^{{{\text{III}}}}\) with \(a_{2}^{{{\text{III}}}}\) and \(t_{2}^{{{\text{III}}}}\) , as shown in Fig. 2 a, where \(a_{1}{\prime} < a_{2}{\prime} < a_{3}{\prime}\) , the second is to increase \(t_{1}^{{{\text{III}}}}\) when \(a_{1}^{{{\text{III}}}}\) is constant, and analyze the variation of the \(E_{{\text{b - j}}}^{{{\text{III}}}}\) and the \(q_{{{\text{loss}}}}^{{{\text{III}}}}\) with \(a_{2}^{{{\text{III}}}}\) and \(t_{2}^{{{\text{III}}}}\) , as shown in Fig. 2 b, where \(t_{1}{\prime} < t_{2}{\prime} < t_{3}{\prime}\) .

The vehicle velocity change with three acceleration values.

Increasing the first acceleration

Firstly, analyze the first scenario, which is when the EV accelerates with three accelerations, the first acceleration \(a_{1}^{{{\text{III}}}}\) is increased when the first acceleration time \(t_{1}^{{{\text{III}}}}\) is constant, and the variation of the \(E_{{\text{b - j}}}^{{{\text{III}}}}\) with the \(a_{2}^{{{\text{III}}}}\) and the \(t_{2}^{{{\text{III}}}}\) is shown in Figs. 3 , 4 and 5 .

The \(E_{{\text{b - j}}}^{{{\text{III}}}}\) change under neighborhood condition (0–40 km/h) with three acceleration values.

The \(E_{{\text{b - j}}}^{{{\text{III}}}}\) change under urban condition (0–80 km/h) with three acceleration values.

The \(E_{{\text{b - j}}}^{{{\text{III}}}}\) change under highway condition (0–120 km/h) with three acceleration values.

For the purpose of further analyzing the relationship between the \(E_{{\text{b - j}}}^{{{\text{III}}}}\) and the \(a_{2}^{{{\text{III}}}}\) and the \(t_{2}^{{{\text{III}}}}\) , and for comparative analysis, a certain boundary is intercepted from Figs. 3 , 4 and 5 for analysis, as shown in Fig. 6 .

The \(E_{{\text{b - j}}}^{{{\text{III}}}}\) change under different working conditions with three acceleration values.

As shown in Figs. 3 , 4 and 5 , when the EV accelerates with three acceleration values, for the neighborhood condition, the \(E_{{{\text{b}} - {\text{j}}}}^{{{\text{III}}}}\) gradually decreases as the increase of \(a_{2}^{{{\text{III}}}}\) and \(t_{2}^{{{\text{III}}}}\) . However, for the urban condition and highway condition, the \(E_{{{\text{b}} - {\text{j}}}}^{{{\text{III}}}}\) decreases first and then increases as the increase of \(a_{2}^{{{\text{III}}}}\) and \(t_{2}^{{{\text{III}}}}\) , and the higher the velocity, the greater the \(E_{{{\text{b}} - {\text{j}}}}^{{{\text{III}}}}\) , and the more obvious this change trend. This indicates that under a given operating condition, when the EV accelerates with three acceleration curve, and \(a_{1}^{{{\text{III}}}} ,t_{1}^{{{\text{III}}}}\) are certain, there is an optimal point ( \(a_{2}^{{{\text{III}}}}\) , \(t_{2}^{{{\text{III}}}}\) ) where \(E_{{{\text{b}} - {\text{j}}}}^{{{\text{III}}}}\) is minimized.

In addition, from Fig. 6 , it can be observed that under the same working condition, when the \(t_{1}^{{{\text{III}}}}\) remains constant, the \(E_{{{\text{b}} - {\text{j}}}}^{{{\text{III}}}}\) decreases as the \(a_{1}^{{{\text{III}}}}\) increases. This indicates that when the EV accelerates with three acceleration values, and when the \(t_{1}^{{{\text{III}}}}\) is constant, the bigger the \(a_{1}^{{{\text{III}}}}\) , the smaller the \(E_{{{\text{b}} - {\text{j}}}}^{{{\text{III}}}}\) . Therefore, when the EV accelerates with multiple acceleration curves, the EV energy consumption can be reduced by appropriately increasing the first segment acceleration.

When the EV accelerates with a convex acceleration curve of three accelerations values, the \(a_{1}^{{{\text{III}}}}\) is increased when the \(t_{1}^{{{\text{III}}}}\) is constant, the relationship between \(q_{{{\text{loss}}}}^{{{\text{III}}}}\) and \(a_{2}^{{{\text{III}}}}\) and \(t_{2}^{{{\text{III}}}}\) is shown in Figs. 7 , 8 and 9 .

The \(q_{{{\text{loss}}}}^{{{\text{III}}}}\) change under neighborhood condition (0–40 km/h) with three acceleration values.

The \(q_{{{\text{loss}}}}^{{{\text{III}}}}\) change under urban condition (0–80 km/h) with three acceleration values.

The \(q_{{{\text{loss}}}}^{{{\text{III}}}}\) change under highway condition (0–120 km/h) with three acceleration values.

Similarly, in order to more intuitively analyze the relationship between \(q_{{{\text{loss}}}}^{{{\text{III}}}}\) and \(a_{2}^{{{\text{III}}}}\) and \(t_{2}^{{{\text{III}}}}\) under different working conditions, the boundary data in Figs. 7 , 8 and 9 are intercepted to analyze, as shown in Fig. 10 .

The \(q_{{{\text{loss}}}}^{{{\text{III}}}}\) change under different working conditions with three acceleration values.

From Figs. 7 , 8 and 9 , it can be found that when the EV accelerates with the convex curve of three accelerations values, under three acceleration conditions of neighborhood, urban, and highway, its \(q_{{{\text{loss}}}}^{{{\text{III}}}}\) gradually decreases with the increase of the \(t_{2}^{{{\text{III}}}}\) , and the \(q_{{{\text{loss}}}}^{{{\text{III}}}}\) increases first and then decreases with the increase of the \(a_{2}^{{{\text{III}}}}\) . Moreover, the lower the vehicle velocity, the more obvious this change trend is, and the higher the vehicle velocity, the smaller the \(q_{{{\text{loss}}}}^{{{\text{III}}}}\) .

In addition, from Fig. 10 , it can also be found that under the same acceleration condition, when the \(t_{1}^{{{\text{III}}}}\) remains constant and increases the \(a_{1}^{{{\text{III}}}}\) , the variation of \(q_{{{\text{loss}}}}^{{{\text{III}}}}\) is different. As shown in the figure, under the same \(a_{2}^{{{\text{III}}}}\) , the larger \(a_{1}^{{{\text{III}}}}\) , the greater \(q_{{{\text{loss}}}}^{{{\text{III}}}}\) ; And under the same \(t_{2}^{{{\text{III}}}}\) , the larger \(a_{1}^{{{\text{III}}}}\) , the smaller \(q_{{{\text{loss}}}}^{{{\text{III}}}}\) , which is different from the variation trend of the \(E_{{\text{b - j}}}^{{{\text{III}}}}\) .with \(a_{2}^{{{\text{III}}}}\) and \(t_{2}^{{{\text{III}}}}\) .

Increasing the first acceleration time

And then analyze the second scenario, that is, when the EV accelerates with a convex acceleration curve of three accelerations values, the \(t_{1}^{{{\text{III}}}}\) is increased when the \(a_{1}^{{{\text{III}}}}\) is constant, and the variation of \(E_{{\text{b - j}}}^{{{\text{III}}}}\) and \(q_{{{\text{loss}}}}^{{{\text{III}}}}\) with \(a_{2}^{{{\text{III}}}}\) and \(t_{2}^{{{\text{III}}}}\) is analyzed. Based on the analysis in “ Increasing the first acceleration ”, in order to further analyze the relationship between \(E_{{\text{b - j}}}^{{{\text{III}}}}\) and \(q_{{{\text{loss}}}}^{{{\text{III}}}}\) and \(a_{2}^{{{\text{III}}}}\) and \(t_{2}^{{{\text{III}}}}\) , as well as compare and analyze the changes in \(E_{{\text{b - j}}}^{{{\text{III}}}}\) and \(q_{{{\text{loss}}}}^{{{\text{III}}}}\) under different acceleration conditions, this section directly intercepts its boundary data for analysis.

As shown in Fig. 11 , the variation of \(E_{{{\text{b}} - {\text{j}}}}^{{{\text{III}}}}\) with \(a_{2}^{{{\text{III}}}}\) and \(t_{2}^{{{\text{III}}}}\) under different acceleration conditions are shown.

From the overall variation trend in Fig. 11 , it can be seen that under neighborhood conditions, the \(E_{{{\text{b}} - {\text{j}}}}^{{{\text{III}}}}\) decreases with the increase of \(a_{2}^{{{\text{III}}}}\) and \(t_{2}^{{{\text{III}}}}\) . However, under urban and highway conditions, the \(E_{{{\text{b}} - {\text{j}}}}^{{{\text{III}}}}\) shows a trend of first decreasing and then increasing with the increase of \(a_{2}^{{{\text{III}}}}\) and \(t_{2}^{{{\text{III}}}}\) , and the higher the vehicle velocity, the more obvious the \(E_{{{\text{b}} - {\text{j}}}}^{{{\text{III}}}}\) change. This indicates that under a given acceleration condition, when an EV accelerates with a three acceleration curve and \(a_{1}^{{{\text{III}}}} ,t_{1}^{{{\text{III}}}}\) is constant, there is an optimal \(a_{2}^{{{\text{III}}}} ,t_{2}^{{{\text{III}}}}\) that where the \(E_{{{\text{b}} - {\text{j}}}}^{{{\text{III}}}}\) is minimized.

Besides, it can also be observed that under the same acceleration condition, when \(a_{1}^{{{\text{III}}}}\) remains unchanged, the \(E_{{{\text{b}} - {\text{j}}}}^{{{\text{III}}}}\) shows a decreasing trend with the increase of \(t_{1}^{{{\text{III}}}}\) . This indicates that when the electric vehicle accelerates with three accelerations, moreover, the \(a_{1}^{{{\text{III}}}}\) is constant, the longer the \(t_{1}^{{{\text{III}}}}\) , the smaller the \(E_{{{\text{b}} - {\text{j}}}}^{{{\text{III}}}}\) . Therefore, when the EV accelerates with multiple acceleration curves, its energy consumption can be reduced by appropriately extending the acceleration time of the first segment.

And the relationship between the \(q_{{{\text{loss}}}}^{{{\text{III}}}}\) and \(a_{2}^{{{\text{III}}}}\) and \(t_{2}^{{{\text{III}}}}\) is shown in Fig. 12 .

As can be seen from Fig. 12 , when the EV accelerates with a convex acceleration curve of three accelerations values, under three acceleration conditions of neighborhood, urban and highway, its \(q_{{{\text{loss}}}}^{{{\text{III}}}}\) decreases with the increase of \(t_{2}^{{{\text{III}}}}\) , and presents a trend of first increasing and then decreasing with the increase of \(a_{2}^{{{\text{III}}}}\) . The lower the vehicle velocity, the more obvious the change trend is, and the higher the vehicle velocity, the smaller the \(q_{{{\text{loss}}}}^{{{\text{III}}}}\) . In addition, it can also be seen that under the same acceleration condition, when the \(a_{1}^{{{\text{III}}}}\) remains unchanged, the \(q_{{{\text{loss}}}}^{{{\text{III}}}}\) increases as the \(t_{1}^{{{\text{III}}}}\) increases, which is exactly contrary to the variation trend of the \(E_{{{\text{b}} - {\text{j}}}}^{{{\text{III}}}}\) .

In summary, when the EV accelerates with different multiple acceleration curves, the changes in energy consumption per kilometer and percentage of battery capacity loss per kilometer with acceleration and acceleration time are different, and the changes in the two are basically opposite. For instance, when the EV accelerates with a three acceleration curve, increasing \(a_{1}^{{{\text{III}}}}\) and \(t_{1}^{{{\text{III}}}}\) appropriately may help reduce \(E_{{{\text{b}} - {\text{j}}}}^{{{\text{III}}}}\) , but \(q_{{{\text{loss}}}}^{{{\text{III}}}}\) will increase. Therefore, the two are contradictory, so in the study of the acceleration process of electric vehicles, it is not only considered energy consumption, but also the impact of the magnitude and number of accelerations in the multiple acceleration curves on the power battery capacity attenuation characteristics.

When the EV accelerates with different acceleration curves, its energy consumption and battery life are different. In this paper, the interaction mechanism between the EV energy consumption and the battery capacity loss under different multiple accelerations curves is studied, and when the EV accelerates with neighborhood, urban, highway conditions, the relationship between the energy consumption, battery capacity attenuation, acceleration and acceleration time is analyzed. The research results indicate that when the EV accelerates with different multiple accelerations curves, their energy consumption and the battery capacity attenuation have different relationships with acceleration and acceleration time, and their variation patterns are basically opposite. Under neighborhood conditions, the \(E_{{{\text{b}} - {\text{j}}}}^{{{\text{III}}}}\) decreases with the increase of \(a_{2}^{{{\text{III}}}}\) and \(t_{2}^{{{\text{III}}}}\) ; and under urban and highway conditions, the \(E_{{{\text{b}} - {\text{j}}}}^{{{\text{III}}}}\) shows a trend of first decreasing and then increasing with the increase of \(a_{2}^{{{\text{III}}}}\) and \(t_{2}^{{{\text{III}}}}\) , and the higher the vehicle velocity, the greater the \(E_{{{\text{b}} - {\text{j}}}}^{{{\text{III}}}}\) , and the more obvious this change trend. However, under three acceleration conditions, the \(q_{{{\text{loss}}}}^{{{\text{III}}}}\) decreases with the increase of \(t_{2}^{{{\text{III}}}}\) , and presents a trend of first increasing and then decreasing with the increase of \(a_{2}^{{{\text{III}}}}\) . The lower the vehicle velocity, the more obvious this change trend, and the higher the vehicle velocity, the smaller the \(q_{{{\text{loss}}}}^{{{\text{III}}}}\) . Thus it can be seen that when the EV accelerates with a three accelerations curve, increasing the first acceleration and the first acceleration time appropriately can help reduce the energy consumption per kilometer, but the percentage of battery capacity loss per kilometer will increase. Therefore, the two are contradictory, so in the subsequent optimization of the acceleration process, not only energy consumption should be considered, but also the impact of the acceleration magnitude, the number of acceleration and acceleration time during acceleration process on the power battery life.

Abbreviations

Traction force (N)

EV velocity (km/h)

Efficiency of the transmission system

Efficiency of the inverter

Efficiency of the electric motor

EV energy consumption per kilometer in the \(l\) acceleration (Wh/km)

Acceleration (m/s 2 )

Percentage of battery capacity loss (%)

Pre-exponential factor

Cumulative ampere-hours ( \({\text{A}}\;{\text{h}}\) )

Gas constant ( \({\text{J/}}\left( {{\text{mol}}\;{\text{K}}} \right)\) )

Percentage of battery capacity loss in the \(l\) acceleration (%)

Percentage of battery capacity loss per kilometer (%/km)

Shabbir, W. & Evangelou, S. A. Real-time control strategy to maximize hybrid electric vehicle powertrain efficiency. Appl. Energy 135 (3), 512–522 (2014).

Article   ADS   Google Scholar  

Hu, J. et al. Parameter matching and optimal energy management for a novel dual-motor multi-modes powertrain system. Mech. Syst. Signal Process. 116 , 113–128 (2019).

Zhu, B. et al. Gear shift schedule design for multi-speed pure electric vehicles. Proc. Inst. Mech. Eng. Part D J. Automob. Eng. 229 (1), 70–82 (2015).

Article   Google Scholar  

Yi, C. et al. Modeling, control, and performance of a novel architecture of hybrid electric powertrain system. Appl. Energy 178 , 454–467 (2016).

Ding, X. et al. A new strategy of efficiency enhancement for traction systems in electric vehicles. Appl. Energy 205 , 880–891 (2017).

Sung, W., Shin, J. & Jeong, Y. Energy-efficient and robust control for high-performance induction motor drive with an application in electric vehicles. IEEE Trans. Veh. Technol. 61 (8), 3394–3405 (2012).

Xinghua, Z., Zhenxing, S. & Deming, W. Optimal efficiency control of direct torque controlled induction motor drives for electric vehicles. Trans. China Electrotech. Soc. 21 (2), 1–5 (2013).

Google Scholar  

Li, S. G. et al. Energy and battery management of a plug-in series hybrid electric vehicle using fuzzy logic. IEEE Trans. Veh. Technol. 60 (8), 3571–3585 (2011).

Chen, Z. et al. Energy management for a power-split plug-in hybrid electric vehicle based on dynamic programming and neural networks. IEEE Trans. Veh. Technol. 63 (4), 1567–1580 (2014).

Jalalmaab, M., & Azad, N. L. A stochastic power management strategy with skid avoidance for improving energy efficiency of in-wheel motor electric vehicles. In Proceedings of the Institution of Mechanical Engineers, Part D: Journal of Automobile Engineering (2018).

Zeng, X. & Wang, J. A two-level stochastic approach to optimize the energy management strategy for fixed-route hybrid electric vehicles. Mechatronics 38 , 93–102 (2016).

Wu, X. et al. Electric vehicles’ energy consumption measurement and estimation. Transport. Res. Part D 34 (34), 52–67 (2015).

Article   CAS   Google Scholar  

Graba, M. et al. Impact of the acceleration intensity of a passenger car in a road test on energy consumption. Energy 25 , 120429 (2021).

Nandi, A. K., Chakraborty, D. & Vaz, W. Design of a comfortable optimal driving strategy for electric vehicles using multi-objective optimization. J. Power Sources 283 , 1–18 (2015).

Article   ADS   CAS   Google Scholar  

Yuan, X. et al. Energy and environmental impact of battery electric vehicle range in China. Appl. Energy 157 , 75–84 (2015).

Xie, Y. et al. Microsimulation of electric vehicle energy consumption and driving range. Appl. Energy 20 , 267 (2020).

He, H. & Cao, J. Energy optimization of electric vehicle’s acceleration process based on reinforcement learning. J. Clean. Prod. 248 , 119302 (2019).

Ge, M. F. et al. A review on state of health estimations and remaining useful life prognostics of lithium-ion batteries. Measurement 174 , 109057 (2021).

Liu, K. Data-driven health estimation and lifetime prediction of lithium-ion batteries: A review. Renew. Sustain. Energy Rev. 20 , 113 (2019).

Song, K. et al. Remaining life prediction of lithium-ion batteries based on health management: A review. J. Energy Stor. 220 , 25 (2023).

Deng, H. W. S. A rest-time-based prognostic model for remaining useful life prediction of lithium-ion battery. Neural Comput. Appl. 33 , 6 (2021).

Lui, Y. H. et al. Physics-based prognostics of implantable-grade lithium-ion battery for remaining useful life prediction. J. Power Sources 485 (6273), 229327 (2021).

Russell, M. B. et al. Stochastic modeling for tracking and prediction of gradual and transient battery performance degradation. J. Manuf. Syst. 20 , 59 (2021).

Xiong, S. et al. Research on capacity characteristics and prediction method of electric vehicle lithium-ion batteries under time-varying operating conditions. J. Energy Stor. 58 , 106334 (2023).

Wang, Y. et al. Remaining useful life prediction of lithium-ion battery based on discrete wavelet transform. Energy Proced. 20 , 20 (2017).

Ahwiadi, M. & Wang, W. An enhanced particle filter technology for battery system state estimation and RUL prediction. Measurement 191 , 191 (2022).

Duan, B. et al. Remaining useful life prediction of lithium-on battery based on extended Kalman particle filter. Int. J. Energy Res. 44 , 3 (2020).

Wei, G. & He, M. An optimal relevance vector machine with a modified degradation model for remaining useful lifetime prediction of lithium-ion batteries. Appl. Soft Comput. 20 , 124 (2022).

Wang, F. K. & Mamo, T. A hybrid model based on support vector regression and differential evolution for remaining useful lifetime prediction of lithium-ion batteries. J. Power Sources 25 , 401 (2018).

Yun, Z. et al. A novel based-performance degradation Wiener process model for real-time reliability evaluation of lithium-ion battery. J. Energy Stor. 50 , 104313 (2022).

Wang, F. K. et al. A hybrid method for online cycle life prediction of lithium-ion batteries. Int. J. Energy Res. 7 , 46 (2022).

Yu, Z. S. Automobile Theory 5th edn. (China Machine Press, 2009).

Wang, J. et al. Cycle-life model for graphite-LiFePO4 cells. J. Power Sources 196 (8), 3942–3948 (2011).

Luo, Y. et al. A study on the driving-cycle-based life model for LiFePO4 battery. Autom. Eng. 37 (08), 881–885 (2015).

Li, L. & Liu, Q. Acceleration curve optimization for electric vehicle based on energy consumption and battery life. Energy 169 , 1039–1053 (2019).

Download references

This research was funded by the Guangdong Basic and Applied Basic Research Foundation (No. 2022A1515110683) and Guangdong Basic and Applied Basic Research Foundation (No. 2022A1515110187).

Author information

Authors and affiliations.

School of Mechatronic Engineering and Automation, Foshan University, Foshan, 528225, China

Qin Liu, Zhongbo Zhang & Jingjing Zhang

You can also search for this author in PubMed   Google Scholar

Contributions

Conceptualization, Q.L.; methodology, Q.L.; formal analysis, Q.L.; investigation, Q.L., Z.Z. and J.Z.; data curation, J.Z..; writing-original draft preparation, Q.L.; writing-review and editing, Z.Z. and J.Z.; supervision, Z.Z.; funding acquisition, Q.L., and Z.Z. All authors have read and agreed to the published version of the manuscript.

Corresponding author

Correspondence to Jingjing Zhang .

Ethics declarations

Competing interests.

The authors declare no competing interests.

Additional information

Publisher's note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ .

Reprints and permissions

About this article

Cite this article.

Liu, Q., Zhang, Z. & Zhang, J. Research on the interaction between energy consumption and power battery life during electric vehicle acceleration. Sci Rep 14 , 157 (2024). https://doi.org/10.1038/s41598-023-50419-3

Download citation

Received : 23 August 2023

Accepted : 19 December 2023

Published : 02 January 2024

DOI : https://doi.org/10.1038/s41598-023-50419-3

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

By submitting a comment you agree to abide by our Terms and Community Guidelines . If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.

Quick links

• Explore articles by subject
• Guide to authors
• Editorial policies

Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

IEEE Account

• Change Username/Password
• Update Address

Purchase Details

• Payment Options
• Order History
• View Purchased Documents

Profile Information

• Communications Preferences
• Profession and Education
• Technical Interests
• US & Canada: +1 800 678 4333
• Worldwide: +1 732 981 0060
• Contact & Support
• About IEEE Xplore
• Accessibility
• Terms of Use
• Nondiscrimination Policy
• Privacy & Opting Out of Cookies

A not-for-profit organization, IEEE is the world's largest technical professional organization dedicated to advancing technology for the benefit of humanity. © Copyright 2024 IEEE - All rights reserved. Use of this web site signifies your agreement to the terms and conditions.

Strategies for Reducing Power Consumption and Increasing Reliability in IoT

• Conference paper
• Open Access
• First Online: 20 March 2019
• Cite this conference paper

You have full access to this open access conference paper

• Ricardo Reis 17  

Part of the book series: IFIP Advances in Information and Communication Technology ((IFIPAICT,volume 548))

Included in the following conference series:

• IFIP International Internet of Things Conference

14k Accesses

3 Citations

2 Altmetric

The Internet of Things (IoT) demands new challenges in the design of computing and electronics components. One of the challenges is the power reduction of this expanding network of connected devices, where the majority is permanently connected. In a large set of applications, another significant issue is reliability, especially on critical areas as health and transport. This paper shows an overview of design strategies that we have developed to reduce power consumption and to increase reliability in circuits that are components of the IoT, as the reduction of the number of transistors in IoT devices, using optimisation techniques and the physical design of circuits tolerant to radiation effects.

You have full access to this open access chapter,  Download conference paper PDF

Similar content being viewed by others

Challenges in the Design of Integrated Systems for IoT

Dependability for the internet of things—from dependable networking in harsh environments to a holistic view on dependability.

A Novel Strategy for Energy Optimal Designs of IoT and WSNs

• Internet-of-Things
• Optimization
• Physical design
• Fault tolerance
• Radiation effects
• Nanoelectronics

1 Introduction

The growing number of connected devices in the Internet of Things (IoT) is one of the reasons for the ever increasing increase in the number of transistors produced annually in the world. Figure  1 , based on (SIA 2005), shows the number of transistors manufactured annually in the world, year by year. This impressive growth is due to 3 main factors: the increasing number of transistors integrated into a chip, the growing number of products that include embedded chips and the increasing number of manufactured copies of each product. The manufacturing cost of a transistor is relatively cheap. In (The Economist 2010 ) a comparison is presented between the cost of a grain of rice and the cost of a transistor. The price of a rice grain can be equivalent to the manufacturing cost of more than 125,000 transistors. This would indicate that there is no need to optimise the number of transistors in a design, since the cost of them is relatively small. But the cost of energy required for the operation of a transistor is increasing a lot. We also have to consider that a high-power consumption can reduce the lifetime of a system, as well as increase the effects of variability that can cause an integrated system to malfunction and/or also reduce its useful life. With the increasing connection of electronic and computational devices on the Internet, that is, in the Internet of Things, power consumption problems tend to get worse, and a lot. How much Power Plants we will need to cope with the IoT/IoE (Internet of Everything) world? This is a major issue.

(adapted from SIA 2005)

Number of transistors produced annually in the world

So, an essential keyword on the Internet of Things is optimisation , especially the optimisation of power consumption, which must be addressed at all levels of abstraction in the design flow of a computer or electronic system. The total power optimisation is a summation of the optimisation done at each level of design abstraction. So, sustainable computing requires optimisation at all design levels of a computer or electronic system design.

2 Internet of Things

The term Internet of Things has already given rise to several other associated terms, such as the Internet of Health (IoH), Internet of People (IoP) and the Internet of Everything (IoE). In fact, the latter term becomes the most comprehensive, but each one of the others has some specific characteristics. When talking about the Internet of Health, which includes real-time monitoring of a person’s clinical conditions, as well as chips injected in a person, the issue of reliability is a key one. And reliability is also related to power consumption in most cases. High power consumption can reduce the lifetime of a system. When it comes to the Internet of People, the issue of people’s security and privacy is of great relevance. But in all cases, the importance of optimising energy consumption is growing more and more.

When considering optimisation, it means that integrated systems must increasingly be dedicated to the intended application to optimise the number of components, that means the number of transistors. Another important strategy for optimisation is the hardware and software codesign, where one can manage the compromise between performance, consumption, and reliability.

Devices connected to the Internet of Things (or the Internet of Everything), can have very different complexities. If it is analysed the complexity considering the number of components, we can find small devices with few transistors and large devices with billions of transistors. Of course, large devices will consume much more power, but we have to consider that most devices on the Internet of Things are devices with a low number of transistors. But, because they are found in large quantities, they can represent a total consumption more important than the consumption of the so-called large devices that are present in a lower number. Therefore, consumption optimisation must be performed on both large and small devices that are present in large quantities. Another aspect to consider is that some devices require the application of reliability techniques (such as those related to transport or health systems), which can increase the number of components, while other devices are not critical, such as a camera or video, where an error in viewing a pixel of an image does not cause significant problems.

Also, we can expect that many systems connected to the Internet of Everything will be Cyber Physical Systems (CPS), that are systems composed by different classes of components like electronic elements, mechanical elements, optical elements, physical sensors, chemical sensors, organic components, and many others. So, it is needed to obtain EDA tools to cope with the design of CPS composed of all these classes of devices.

Figure  2 (The Connectivist 2014 ) shows an estimate of the number of devices connected to the Internet since 1992 when they were about 1 million devices. By 2020 when it is estimated that there will be more than 50 billion devices connected in the network, and there are currently around 35 billion connected devices. In (IHS Markit 2018 ) the number of devices connected to the network in 2018 is shown by industrial and commercial sectors, where almost half is in the area of communication. The significant growth in the number of connected devices to the Internet has naturally led to a considerable increase in the energy consumed in the Internet of Things. For how long will we have the energy to meet this growing demand? Therefore, it is necessary to use techniques to minimise the energy consumption of each connected device in the Internet of Things.

(adapted from The Connectivist 2014 )

Number of devices connected on the Internet

The Internet of Health (IoH) is a significant way to increase the life of human beings but also to improve life quality. Some of the examples of devices to be connected to the IoH are: Glasses that can advise eye correction; Toothbrush that can find cavities and breath issues; Razor that identify acne; Pacemakers that broadcast data to cardiologist; Underwearables that can provide early detection of cancer and other anomalies; Combs that can scan for fungus and hair loss; Earphones that does measurement of hearing, analysis of emotional level; Watches able to measure parameters like blood pressure, heart rate and others.

In critical areas such as the design of implanted devices (chips) in humans (Fig.  3 ), the reliability of the implanted systems is obviously critical. Some of the techniques used are based on the triplication of circuits and the temporal analysis of the propagation of a signal. Previously, the design of fault-tolerant circuits, to cope with radiation effects, was mainly in circuits that were sent to space. With the reduction of the value of the supply voltage of integrated circuits, nowadays the integrated circuits for use at ground level are also sensitive to errors caused by the radiation incident on the earth. Therefore, in critical areas such as implanted chips in humans, it is necessary to implement radiation effects tolerance techniques (Velazco et al. 2007 ). Also, critical systems used on the Internet of Health should be tolerant to any kind of noise (internal or external to the human body). They also must have a larger lifetime as possible, for obvious reasons and also should cope with environmental variability.

The implantation of Chip Systems in humans demands reliability and ultra-low consumption

Also, there is the effect of “ageing”, that is, the ageing of the circuit, which is more eminent in nanometric technologies (Vasquez et al. 2012 ). One of the most important effects is known as NBTI (Negative Bias Temperature Instability) that alters the threshold voltage of the PMOS transistors, degrading the operation of the transistor. Another effect that causes failures in circuits throughout their life is the effect of electromigration, which can cause short circuits or rupture of connections (Fig.  4 ). In order to increase the lifetime of the chips, it is necessary to use physical design techniques that reduce the probability of electromigration (Posser et al. 2016 , 2017 ).

(Geden 2011 ).

Exemplo of a Void (open circuit) and hillock (short circuit)

3 Electronic Design Automation (EDA) Tools

The use of EDA tools is essential for optimising energy consumption and increasing reliability, as the design flow has a large set of steps as well the number of components of a cheap can reach billions off transistors. In Fig.  5 we can see the floorplan of an integrated circuit, where the hotter colours show regions (hot spots) with higher energy consumption, indicating that in some points there is a significant concentration of power consumption. One way to deal with the problem is to modify the placement of the logic cells in the circuit to distribute the cells with the highest energy consumption over the entire circuit area. But this must be done without compromising the area, wirelengh and operating frequency specifications (much depends on the routing). Another way is to decrease the number of transistors, since the static consumption is related to the number of transistors (Reis 2011A ).

View of the consumption density on a chip

4 Power Consumption Reduction by Reducing the Number of Transistors

The reduction of the power consumption of a System on a Chip (SoC) is a function of a sum of techniques and strategies of design applied in different levels of abstraction in the design flow of an integrated system (Reis 2010 ). The summation of the gains is that it will set the total gain in power reduction. When we deal with the physical synthesis of a system on a chip, one technique is the optimisation of the number of components, that is, the number of transistors. In Fig.  6 (Reis 2011A ) we can observe two solutions for the implementation of the same equation. The first solution makes use of 4 basic logic gates (3 NOR 2-input ports and one CMOS inverter), using a total of 14 transistors. The second solution makes use of only one logic gate, which performs the same function but with only 8 transistors. That is, the second solution, having a reduction in the number of transistors, will also have a proportionally smaller static power consumption. Furthermore, in the example of Fig.  6 , we can see that the first solution also has 3 connections between the basic gates (and therefore even vias and contacts) that are eliminated in the second option with only one logic gate.

(Reis 2011A )

Two options for the implementation of the same function

This elimination of connections is increasingly important because it decreases the number of connections to be implemented using the different metal layers. The decrease in the number of connections decreases the density of connections and, therefore, increases the routability of the circuit and also contributes to reduce the average length of the connections, which implies in a reduction of the delay. In modern technologies, the delay in connections is so or more significant than the delay in the switching of logic gates. A greater spacing between the connections also contributes to an increase of reliability, due, for example, to the reduction of the possibility of electromigration, as already mentioned above.

The reduction of the number of transistors depends on the use of efficient Electronic Design Automation (EDA) tools that transform the logical equations of a system so that in addition to mapping equations in CMOS gates, make optimum use of complex logic gates. In (Conceição et al. 2017 ) we present a tool to reduce the number of transistors in a circuit through the fusion of networks of transistors that present fanout equal to 1. Also, it is fundamental the use of an automatic synthesis tool that can perform the automatic layout of any logical function. There is no use to achieve a logical optimisation if it is necessary to map (transform) the equations according to the logic gates available in a traditional cell library [which have few functions, in general, no more than 100 functions], as is still done when using traditional EDA systems. This mapping step is called technology mapping, and it represents a step of deoptimization. With this aim, we have developed automatic layout synthesis tools such as ASTRAN (Ziesemer and Reis 2015 ) (Fig.  7 ), which allows automatic generation of the layout of any network of transistors (Reis 2011A , B ).

(Ziesemer and Reis 2015 )

Transistor network layout generated automatically

Another technique to reduce consumption is through the sizing of the transistors. Modern integrated circuit manufacturing technologies show a significant increase in static power consumption that is often greater than dynamic power consumption. One way to mitigate power consumption, especially the static one, is to carry out a sizing of transistors to optimise power consumption (Posser 2011 ). In (Reimann et al. 2016 ) significant decreases in consumption are obtained through the use of automatic transistor sizing tools. This is also called cell selection, where the cells are selected from a cell library. In this case, cell selection means the selection of cells with a specific size and Vth (threshold voltage). In traditional cell libraries, one function has in general 3 sizings (one for less area, one for less power, and one for less delay) and 3 Vth (threshold voltage).

5 Reliability

As in the reduction of power consumption, in the design of critical systems, it is needed to use techniques to increase reliability at different levels of design abstraction. At the architectural level, a very applied method is the redundancy of modules, especially triple module redundancy (TMR) (Kastensmidt et al. 2006 ). Another is the temporal redundancy (Nicolaidis 1999 ) where a signal traverses two paths, one with higher delay and another one with less delay. The difference of delay must be longer than the duration of a transient. Comparing the signal after traversing the two paths indicates whether there has been a transient propagation or not. At the physical level, we can apply different techniques to reduce or avoid problems such as electromigration (Posser et al. 2016 , 2017 ). In the example of Fig.  8 , the position of the output pin in the centre (point 4) increases the lifetime of the circuit because it allows reducing the maximum density of current in the segments of the metal layer.

(Posser et al. 2016 , 2017 )

Changing current density with the change of position of the output pin

In (Velazco et al. 2007 ) it is presented a series of works aimed at mitigating the effects of radiation on integrated circuits. In (Kastensmidt et al. 2006 ; Neuberger et al. 2014 ; Gennaro et al. 2017 ; Aguiar et al. 2016 ; Lazzari et al. 2011 ) we present some of the results that our research group has obtained in the development of techniques aiming the design tolerant to faults due to transients, as the effects due to radiation.

6 Hardware Accelerators

The evolution of computer architectures, that today means, the evolution of microprocessor architectures has been very significant. In the 1970s, one marketing argument from microprocessor producers was the number of instructions that the microprocessor could execute as well as the clock frequency of the microprocessor. In the last decades, there has been a change of paradigm, discontinuing the race for the increase of the clock frequency, because the increment of the clock means an increase of the dynamic consumption. Instead, there was an increase in the number of cores (CPUs) aiming at increasing performance. Initially with homogeneous cores and later with heterogeneous cores.

Currently, we can find chips with multiple CPUs and several GPUs (as can be seen in Fig.  9 (Shao 2016 ) showing the floorplan of the A8 microprocessor (from Apple). In this same figure, it can be observed that about half of the area is occupied with hardware accelerators, which are modules dedicated to the execution of a specific function. For example, an encryption module placed next to the output/input pins and which will encode the output data and decode the received data. So, the execution of this function will be faster, because it is done by a dedicated module (that means smaller) and with only the needed number of components to perform that function. It also will consume less power.

(AnandTech 2014 ; Shao 2016 )

Apple 8 floorplan with 29 hardware accelerators

A more important fact is that the use of hardware accelerators leads to greater energy efficiency (allowing more sustainable computing), mainly due to the reduction in the number of components used to perform a function. At any given time, only the hardware accelerators in use at that time are being powered. So, the hardware accelerators that are not in use are disconnected from the power supply. This strategy is also known as “Dark Silicon”. We can even predict architectures consisting essentially of hardware accelerators, with only one or two small CPUs to manage these hardware accelerators.

The introduction of an NPU in A11 is another element characterising the heterogeneity of the SoC (chip system). And we can expect increasingly heterogeneous architectures, with dedicated modules for different operations to be performed by a SoC. In Fig.  10 (Techinsights 2017 ) the floorplan of the Apple A11 microprocessor is presented, where one of the modules is an NPU (Neural Processing Unit). NPU is mostly dedicated to facial recognition (Techinsights 2017 ), processing machine learning tasks more efficiently, consuming less energy than CPUs do. The CPUs occupy about 15% of the area of the chip and 6 GPUs occupy about 20% of the area. Most of the area is filled with the hardware accelerators. That is, it is growing in the architecture of Apple microprocessors the use of hardware accelerators.

(Techinsights 2017 )

Apple 11 floorplan with an NPU

7 Conclusions

To have sustainable computing, when the number of connected devices in the Internet of Things is fast increasing, it is fundamental the design of devices optimised regarding energy consumption. Most of the chips produced today use much more transistors than necessary to perform a function. So, there is a significant space for the optimisation of the number of components. In many devices related to critical applications, the application of techniques for fault tolerance is also fundamental, as nowadays circuits at ground level can have faults due to radiation effects. The reduction of power consumption must be treated at all design abstraction levels in a synthesis flow of integrated systems, from the specification of them in high-level languages to the physical synthesis. It was presented several works that were developed to reduce the power consumption and increase the reliability of integrated systems on a chip, and more details are shown in the mentioned references. The keyword in the age of the Internet of Things is optimisation .

Aguiar, Y., Zimpeck, A., Meinhardt, C., Reis, R.: Permanent and single event transient faults reliability evaluation EDA Tool. In: Microelectronics Reliability, September 2016, vol. 64, pp. 63–67. Elsevier B.V., Amsterdam (2016). ISSN 0026-2714

Google Scholar  

AnandTech (2014). https://www.anandtech.com/show/8562/chipworks-a8

Conceição, C., Moura, G., Pisoni, F., Reis, R.: A cell clustering technique to reduce transistor count. In: 24th IEEE International Conference on Electronics, Circuits and Systems – ICECS 2017, Batumi, Georgia, 5–8 December 2017, pp. 186–189 (2017). https://doi.org/10.1109/icecs.2017.8291996

Gennaro, R., Rosa, F., Oliveira, A., Kastensmidt, F., Ost, L., Reis, R.: Analyzing the impact of fault tolerance methods in ARM processors under soft errors running Linux and parallelization APIs. IEEE Trans. Nucl. Sci 64 (8) (2017). https://doi.org/10.1109/tns.2017.2706519 . ISSN 1558–1578

Geden, B.: Understand and avoid electromigration (EM) & IR-drop in custom IP blocks. Synopsys (2011)

Lazzari, C., Wirth, G., Kastensmidt, F., Anghel, L., Reis, R.: Asymmetric transistor sizing targeting radiation-hardened circuits. J. Electr. Eng. (2011A). https://doi.org/10.1007/s00202-011-0212-8 . Accessed June 2011

Kastensmidt, F., Carro, L., Reis, R.: Fault-Tolerance Techniques for SRAM-Based FPGA, pp. 1–183. Springer, New York (2006). https://doi.org/10.1007/978-0-387-31069-5 . ISBN 0-387-31068-1

Book   Google Scholar  

Neuberger, G., Wirth, G., Reis, R.: Protecting Chips Against Hold Time Violations Due to Variability, pp. 1–107. Springer, New York (2014). https://doi.org/10.1007/978-94-007-2427-3 . ISBN 978-94-007-2426-6

Nicolaidis, M.: Time redundancy based soft-error tolerance to rescue nanometer technologies. In: Proceedings of IEEE VLSI Test Symposium, vol. 17, pp. 86–94. IEEE Computer Society (1999)

Posser, G., Flach, G., Wilke, G., Reis, R.: Gate sizing minimizing delay and area. In: ISVLSI 2011, IEEE Computer Society Annual Symposium on VLSI, Chennai, India, 4–6 July 2011, pp. 315–316 (2011). https://doi.org/10.1109/isvlsi.2011.92 . ISBN 978-0-7695-4447-2

Posser, G., Mishra, V., Jain, P., Reis, R., Sapatnekar, S.: Cell-internal electromigration: analysis and pin placement based optimization. IEEE Trans. Comput.-Aided Des. Integr. Circ. Syst. 35 (2), 220–231 (2016). https://doi.org/10.1109/TCAD.2015.2456427 . ISSN: 0278-0070

Article   Google Scholar  

Posser, G., Sapatnekar, S., Reis, R.: Electromigration Inside Logic Cells, 118 p. Springer (2017). https://doi.org/10.1007/978-3-319-48899-8 . ISBN: 978-3-319-48898-1

Reimann, T., Sze, C., Reis, R.: Challenges of cell selection algorithms in industrial high performance microprocessor designs. In: Integration, vol. 52, pp. 347–354. Elsevier B.V., January 2016. https://doi.org/10.1016/j.vlsi.2015.09.001 . ISSN 0167-9260

Reis, R.: Redução de Consumo pela Otimização de Componentes. In: SEMISH 2010, Anais do 37º Seminário Integrado de Software e Hardware, Belo Horizonte, 21 a 22 de julho de 2010, pp. 371–379 (2010). ISSN 2175-2761

Reis, R.: Design automation of transistor networks, a new challenge. In: IEEE International Symposium on Circuits and Systems, ISCAS 2011, Rio de Janeiro, Brasil, 15–19 May 2011, pp. 2485–2488. IEEE Press (2011A). https://doi.org/10.1109/iscas.2011.5938108 . ISBN 978-1-4244-9472-9

Reis, R.: Power consumption & reliability in NanoCMOS. In: IEEE NANO, 11th International Conference on Nanotechnology, Portland, USA, 15–19 August 2011 (invited talk), pp. 711–714 (2011B). https://doi.org/10.1109/nano.2011.6144656 . ISBN 978-1-4577-1515-0

The Connectivist (2014). http://ow.ly/i/5vph6/original

The Economist (2010). Accessed 6 Sept 2010

SIA: Semiconductor Industry Association, Rrebooting the IT Revolution (2015). http://www.semiconductors.org/clientuploads/Resources/RITR%20WEB%20version%20FINAL.pdf

IHS Markit: IoT Trend Watch 2018 (2018). https://ihsmarkit.com/forms/thankyou.html?efid=t+m2jEyFYkJQYyoP3YvuHA==&&gasc_id=862037098&&gasc_label=scrXCLnM7m0Q6siGmwM

Techinsights (2017). http://techinsights.com/about-techinsights/overview/blog/apple-iphone-8-teardown/

Vazquez, J., et al.: Delay sensing for long-term variations and defects monitoring in safety–critical applications. Analog Integr. Circ. Sig. Process. 70 (2), 249–263 (2012). https://doi.org/10.1007/s10470-011-9789-0 . ISSN 0925-1030

Velazco, R., Fouillat, P., Reis, R.: Radiation Effects on Embedded Systems. Springer, New York (2007). ISBN 978-1-4020-5645-1

Shao, Y.S.: Design and modeling of specialized architectures. Ph.D. thesis, Harvard, May 2016. https://ysshao.github.io/papers/shao2016-dissertation.pdf

Ziesemer, A., Reis, R.: Physical design automation of transistors network. Microelectron. Eng. 148 , 122–128 (2015). https://doi.org/10.1016/j.mee.2015.10.018 . ISSN 0167-9317

Download references

Acknowledgements

We thank CNPq, FINEP, Fapergs, and CAPES for financial support for the development of our team’s work, as well as the master’s and doctoral students of PGMICRO and PPGC and students of Scientific Initiation who have contributed to the research works that served as the basis for this paper.

Author information

Authors and affiliations.

Instituto de Informática – Universidade Federal do Rio Grande do Sul (UFRGS), Caixa Postal 15.064, 91501-970, Porto Alegre, RS, Brazil

Ricardo Reis

You can also search for this author in PubMed   Google Scholar

Corresponding author

Correspondence to Ricardo Reis .

Editor information

Editors and affiliations.

De Nederlandsche Bank, Amsterdam, The Netherlands

Leon Strous

Google, Reston, VA, USA

Vinton G. Cerf

Rights and permissions

Open Access This chapter is licensed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license and indicate if changes were made.

The images or other third party material in this chapter are included in the chapter's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the chapter's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.

Reprints and permissions

Copyright information

© 2019 The Author(s)

About this paper

Cite this paper.

Reis, R. (2019). Strategies for Reducing Power Consumption and Increasing Reliability in IoT. In: Strous, L., Cerf, V. (eds) Internet of Things. Information Processing in an Increasingly Connected World. IFIPIoT 2018. IFIP Advances in Information and Communication Technology, vol 548. Springer, Cham. https://doi.org/10.1007/978-3-030-15651-0_8

Download citation

DOI : https://doi.org/10.1007/978-3-030-15651-0_8

Published : 20 March 2019

Publisher Name : Springer, Cham

Print ISBN : 978-3-030-15650-3

Online ISBN : 978-3-030-15651-0

eBook Packages : Computer Science Computer Science (R0)

Share this paper

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Publish with us

Policies and ethics

Societies and partnerships

• Find a journal
• Track your research

Advances, Systems and Applications

• Open access
• Published: 17 December 2022

A systematic review on effective energy utilization management strategies in cloud data centers

• Suraj Singh Panwar 1 ,
• M. M. S. Rauthan 1 &
• Varun Barthwal 1  

Journal of Cloud Computing volume  11 , Article number:  95 ( 2022 ) Cite this article

6165 Accesses

10 Citations

Metrics details

Data centers are becoming considerably more significant and energy-intensive due to the exponential growth of cloud computing. Cloud computing allows people to access computer resources on demand. It provides amenities on the pay-as-you-go basis across the data center locations spread over the world. Consequently, cloud data centers consume a lot of electricity and leave a proportional carbon impact on the environment. There is a need to investigate efficient energy-saving approaches to reduce the massive energy usage in cloud servers. This review paper focuses on identifying the research done in the field of energy consumption (EC) using different techniques of machine learning, heuristics, metaheuristics, and statistical methods. Host CPU utilization prediction, underload/overload detection, virtual machine selection, migration, and placement have been performed to manage the resources and achieve efficient energy utilization. In this review, energy savings achieved by different techniques are compared. Many researchers have tried various methods to reduce energy usage and service level agreement violations (SLAV) in cloud data centers. By using the heuristic approach, researchers have saved 5.4% to 90% of energy with their proposed methods compared with the existing methods. Similarly, the metaheuristic approaches reduce energy consumption from 7.68% to 97%, the machine learning methods from 1.6% to 88.5%, and the statistical methods from 5.4% to 84% when compared to the benchmark approaches for a variety of settings and parameters. So, making energy use more efficient could cut down the air pollution, greenhouse gas (GHG) emissions, and even the amount of water needed to make power. The overall outcome of this review work is to understand different methods used by researchers to save energy in cloud data centers.

Introduction

Cloud Computing has become a flexible, resourceful, efficient, and prevalent computational technology that offers users reliable, customized, and dynamic computing environments. Cloud applications are hosted on high-capacity systems and storage devices in multiple locations around the world. Rapid demand for cloud-based facilities essentially requires the development of massive data centers that consume excessive amounts of electricity. Optimization of energy can be proficient by uniting resources based on current utilization, well-organized network, and the thermal position of nodes and computing equipment. Because maximizing the utilization of physical servers is essential in lowering a data center’s (DC) energy demand, virtual machines (VMs) have been effectively introduced in DCs to increase server resource utilization. A method for cost-effective VM migration based on fluctuating electricity prices cuts the energy costs of running a cloud service by a large amount.

Cloud computing is an extension of parallel computing, utility computing, cluster computing, and grid computing. It is distributed in nature, so a group of independent resources are spread in remote locations. Cloud computing is defined by NIST as “a model for enabling ubiquitous, convenient, on-demand network access to a shared pool of configurable computing resources (e.g., storage, networks, servers, services, and applications) that can be rapidly provisioned and released with minimal management effort or service provider interaction” [ 1 , 2 ].

The service models of cloud computing are Software as a Service (SaaS), Platform as a Platform (PaaS), and Infrastructure as a Service (IaaS). In SaaS, the client has access to cloud services via a web browser to maintain user interaction and data in the cloud. PaaS is a service that allows customers to use the platform and tools instead of purchasing and paying for software licences for platforms such as operating systems, databases, and intermediary applications.

IaaS means the necessary environment to facilitate cloud services. It contains the pool of hardware resources related to computing, storage, networking, etc. Based on the model of deployment, clouds are categorized into four types. The term “public cloud” refers to an infrastructure that allows the general public to store and access data from any location using a client device with an internet connection. Private Cloud: A private cloud or enterprise cloud is one where the facilities and infrastructure are available for the organization or partner’s use only. A Hybrid Cloud: When a private cloud is combined with public cloud computing. Community Cloud: Resources are shared by multiple organizations that serve a particular community with common concerns [ 3 , 4 ].

Today, research community’s top priorities are energy conservation and effectiveness. The issue of excessive energy utilization arises as a result of unexpected and rapid changes in the environment around the globe [ 5 ]. The levels of carbon footprint and Green House Gases (GHG) in the environment have rapidly increased. The information and communication technology (ICT) industry has been identified as the primary emitter [ 6 ]. The rise of sophisticated and diverse data-intensive services and applications has exacerbated energy challenges. The intensity and constant growth of ICT energy demand have necessitated not only meeting energy requirements but also developing and implementing efficient energy-savings methods. According to a 2016 survey, the total global energy consumption and CO 2 emissions are expected to rise by 48% and 34%, respectively, between 2010 and 2040 [ 7 ]. Also, the Climate Action Group found that the world released 32 gigatonnes of CO2 in 2015 [ 8 ].

The paper is organized as follows: First, a brief introduction of cloud computing, motivation, virtualization, energy consumption, SLAV, VM consolidation, CloudSim, workload datasets, purpose, and classification of the survey have been explained. Further next section  defines the discussion, analysis, objectives, limitations, and evaluation of existing related work for heuristic, metaheuristic, machine learning, and statistical techniques with tools, performance metrics, and comparisons with their benchmark algorithms related to energy consumption. In last section, result analysis, major challenges, suggestions, and future work are elaborated. Finally, the summary and conclusion of the review paper is summarised to improve energy efficiency in cloud data centers. 

The idea behind cloud computing is to provide on-demand quick access to cloud data centers and to administer the operations from a remote location. Cloud computing operates on a pay-as-you-go pricing model, allowing organizations to reduce operational costs and manage infrastructure more effectively. The motivation for conducting the survey, entitled ‘Effective Energy Utilization Management Strategies in Cloud Data Centers’ is to reduce power utilization in well-organized data centers with the help of VM consolidation. There are several proposed resource management approaches for several computing domains, but only a few addresses the issue of energy efficiency in addition to optimizing profit and service quality. Many magnificent studies have been devoted to confirming the consolidation achieved to an appreciable value, but it is still in its developing stage. Various survey papers on load balancing [ 9 , 10 ], resource provisioning [ 11 ], resource scheduling [ 12 , 13 ], resource allocation [ 14 ], and resource utilization [ 15 ] have been published. These surveys explored resource management classification and compared state-of-the-art algorithms based on many significant characteristics of cloud computing. But the classification and techniques related to effective energy utilization approaches have not been discussed in detail in the current study. As a result, there is a need for a complete and systematic assessment of existing energy-efficient strategies, as well as their limitations, to entice academics to work in this domain. This study provides an attempt to investigate the categorization of energy-efficient virtual machine consolidation thoroughly, which will be useful for future research in developing new energy-efficient algorithms or methodologies. The limitations of current approaches are emphasized to inspire future research work challenges and the development of algorithms. The following are the primary contributions of the review paper:

Investigate and analyze the various existing energy-efficient methods in cloud data centers.

Classification of VM management using heuristics, metaheuristics, machine learning, and statistical techniques.

The most important parts of each classification are explained, and a summary of future research goals is also given.

An overview of the tools and workload traces that can be used in the cloud environment to measure how well an algorithm works has been shown.

Overall, the goal is to ascertain how well computers use their resources and consume the least amount of energy possible while still meeting SLA limits for RAM, CPU, bandwidth, etc.

Virtualization

Virtualization technology manages massive data centers more efficiently by allowing several applications, software, and operating systems to run on a single host. It bridges the hardware resources and the operating system, dividing the cloud services into logical units called virtual machines (VMs) [ 16 ]. Virtualization solutions such as Xen, VMware, and KVM (Kernel-based VM) are used to construct virtual environments in cloud data centers [ 17 ]. Figure  1 displays the classification of energy management techniques.

Classifications of energy management techniques

Energy – efficient cloud computing

Cloud computing offers virtualized resources in cloud data centers for handling several requests for different tasks. A cloud data center’s infrastructure often consists of thousands of huge computing hosts with fast processing resources that use a tremendous amount of energy. So, energy-efficient cloud computing is a step forward in analyzing, and implementing global energy reductions in a system providing quality of services while lowering costs [ 18 ]. We may conserve energy by consolidating hardware and minimising repetition. If necessary, services should be able to be virtualized and controlled within a data centre, as well as relocated to other locations. To support energy efficiency in the future, machine-readable accounting of the requirements and characteristics of applications, networks, servers, or even entire sites must be available [ 19 ]. Energy consumption in a cloud DC organization with m nodes and n switching elements is written as follows [ 20 ].

PUE (Power Usage Effectiveness) is a typical efficiency indicator for data center energy usage that describes how satisfactorily a data center utilizes energy. The PUE formula is well described by eq. ( 2 ), which says that it is the ratio of the total energy used in the building to the total energy used by IT equipment in a data center:

As measured at the meter, the electricity dedicated to the data center facility is included in the total facility energy which includes all loads, such as IT equipment, lighting systems, cooling systems, and power supply components. Total IT equipment includes all the energy used by storage, computing, networking, and other control devices like KVM switches, displays, workstations, and laptops, etc.

Energy consumption and service level agreement

Cloud service providers develop an infrastructure where large numbers of high-end computers or servers are installed and interconnected. This hardware platform provides computing, storage, and different amenities to the customer via the internet. As a cloud service provider, the management of power consumption becomes a crucial task. Effective management of resources are required to optimize power utilization, quality of service, cost-effective, and maximize performance with accuracy. In addition to energy utilization and SLA violation, financial expenses and CO 2 emissions from data center cooling systems have a substantial impact on the environment [ 21 ].

The most significant challenges in cloud computing are task scheduling, resource utilization, load balancing [ 9 ], SLA, quality of service (QoS), scalability, disaster recovery, safety, fault tolerance, resource management, energy efficiency, virtual machine migration, and automated service provisioning [ 22 ]. This review work focuses on the previous study of energy efficiency or power consumption, which should be minimized. However, energy and SLAV are inversely associated, as illustrated in Fig.  2 . There is a trade-off between energy consumption and performance (QoS). Performance is described in terms of SLA, which defines the standards and services with throughput, service time, delay time, and reaction time given by the deployed system. A simulation for the environment mentioned in [ 23 ] is performed using the LrMmt host overload detection method with various safety parameter values. The allocation strategy uses the tuning parameters to anticipate the CPU utilization by the host. For example, if the parameter is set to 1.2, the projected utilization is increased by 20%, providing the host a 20% safety buffer to enhance its consumption without violating SLAs. The results reveal that when this value drops, more VMs are packed into a host. Figure  2 shows that when the safety parameter falls, the EC drops and the number of SLA breaches grows. As a result, the parameters must be set to balance the SLAV and EC.

The output of Energy consumption and SLAV by LrMmt policy

As a result, cloud providers must cope with the trade-off between energy-performance and reducing energy consumption while fulfilling QoS standards. Buyya et al. [ 23 ] showed that when the utilization threshold increases, energy usage is reduced but the percentage of SLAV is also increased. This is because a higher utilization threshold permits more aggressive VM consolidation but at the expense of an increased chance of SLAV. As a result, to save energy, aggressive VM consolidation may result in performance or QoS deterioration, resulting in SLAV. So, while reducing energy utilization, SLAV should also be considered to ensure high adherence to the SLA. To minimize EC and SLAV, the combined metric ESV that captures energy consumption and the level of SLAV is calculated for the performance parameter, as EC decreases with the increased level of SLAV.

VM consolidation

In a cloud data center, a central node routes customer applications to the appropriate servers. This facility is known as VM scheduling. To advance the quality of services and efficient management of power consumption, VM scheduling has been done in such a way that a minimum number of hosts are in a state of running. This method is also known as Dynamic Consolidation of Virtual Machine (DCVM) [ 23 ]. Predicting host utilization is an ongoing research effort, and a variety of solutions have been proposed. A single host can host more than one VM, and as per user request, VMs use hosts’ resources. When the request of resource host is underutilized or overutilized then VM has to be relocated. This action is known as VM migration and is a popular approach for controlling power consumption. Migration of virtual machines from underutilized and overloaded hosts is a difficult job. To shrink the quantity of VM migrations, appropriate VM selection, and VM placement methods must be developed. When a VM moves from a host that is too busy, both the source host and the new host use power without providing any services.

CloudSim [ 24 , 25 ] is free, accessible software for simulating cloud computing services and frameworks. This simulator was designed by the CLOUDS (Cloud Computing and Distributed Systems) research laboratory at Melbourne University. Written entirely in Java, CloudSim is a toolkit used to prototype and imitate a cloud computing setting. It enables the modelling of virtualized environments, as well as their administration and on-demand resource management [ 11 ]. This simulator is also enhanced to allow for energy-aware models and power models to simulate service applications with variable workloads.

Workload data

As CloudSim simulator is the preferred tool for research where the workload traces of data is used to test the algorithm. Many researchers are working on PlanetLab or Bitbrains data workloads, where a file associated with one VM denotes the CPU utilization of physical machines. Some workload traces include dynamic data such as CPU, RAM, disc, and network I/O values [ 26 ]. PlanetLab workload traces [ 27 ] with statistical features are given in Table  9 . Bitbrains is a cloud service agency that focuses on managed hosting and enterprise business computation [ 28 ]. Bitbrains’ dataset comprises resources that are used by 1750 VMs from a distributed cloud center. This dataset is published online in the Grid workloads archive [ 29 ]. It is divided into fastStorage and Rnd traces. The fastStorage contains 1250 VMs, and Rnd traces have 500 VMs. The fastStorage data is divided into one file per VM, with each file comprising 30 days of data collected every 5 minutes. Bitbrains workload traces with statistical features are given in Table 10 Apart from PlanetLab or Bitbrains, some other workload traces such as Google cluster traces [ 30 , 31 ], Alibaba cluster [ 32 ], Azure trace [ 33 ], microservices cluster [ 34 ], etc. are also used by researchers. In May 2011, Google released a 29-day cluster trace – a history of every job request, scheduling choice, and resource use statistics for all tasks in a Google Borg computing cluster. The Alibaba group publishes the Alibaba cluster trace program. Their initiative assisted researchers, students, and others interested in the subject by providing cluster traces from the real-world. This allows a better understanding of the features of current internet data centers (IDCs).

Purpose and classification of survey

The Cloud data centers that host and store data are the backbone of cloud computing, which consists of networked computers, power supply, cables, and other components. Data centers that host cloud applications require a lot of energy for resources, leading to high operational costs and carbon release. As expected from a survey, total global energy utilization and carbon dioxide emissions are expected to rise by 48% and 34%, respectively, between 2010 and 2040 [ 7 ]. According to a McKinsey analysis [ 35 ], “the entire expected energy expense for cloud data centers in 2010 was $11.5 billion, and cost of energy doubles every five years in a typical data center”. So, cloud data centers are becoming very expensive and harmful to the environment. The authors of [ 36 ] has conducted a systematic examination of the present status of software solution that helps in reduction of energy consumption in data centers and also stated the impact of data centers on the environment. In [ 37 ] the use of big data, cloud, and IoT leads to higher demands for hyperscale data centers (HDCs) for data storage and processing. The analysis of 60 regions done by the researchers has predicted the overall increase in the energy consumption of HDCs, carbon emissions and electricity costs, that focus the purpose of the survey.

The main challenge is to set up a balance between system performance and energy utilization [ 38 ]. In this detailed systematic survey, a balance between energy efficiency and performance using VM placement [ 39 ], VM selection, and migrations [ 40 ], has been analyzed for data storage and processing [ 41 ]. This paper analyses the approaches performed by various academicians, organizations, researchers in the field of energy consumption in cloud data centers during VM scheduling. Researchers have also compared their method with the benchmark method using different algorithms of heuristics, metaheuristics, machine learning, and statistical methods. Their results show an improvement in energy-saving and thus reduces power consumption. Fig.  3 shows the detailed classification of effective energy management strategies in cloud data centers categorized into four groups i.e., heuristics, metaheuristics, machine learning and statistical. For efficient energy utilization the stages of dynamic VM consolidation is shown in Fig. 4 .

Classification of Energy Management Strategies in Cloud

Stages of Dynamic VM Consolidation

Related work

In cloud computing effective energy management strategies related work has been provided by researchers. Many researchers have applied different techniques for VM management and energy-efficient strategies to reduce energy consumption in cloud data centers. Some have focused on heuristic methods as classified in Fig.  5 , some on metaheuristics as classified in Fig.  6 and Fig.  7 , some on machine learning as described in Fig.  8 , and others on statistical methods categorized in Fig.  9 . To balance the load and decrease energy usage, cloud data centers use live VM migration [ 42 ]. VMs are dynamically distributed among the hosts during a live migration to reduce the number of low utilization hosts and maximize the number of high utilization hosts. Although dynamic VM consolidation can significantly reduce energy usage, live migration increases service level agreement violations. As a result, in order to decrease energy usage while satisfying service level agreements, cloud data centers require an effective dynamic VM consolidation solution. The dynamic VM consolidation procedure can often be divided into three parts [ 43 ].

Classification of existing research using a Heuristic method

Classification of Meta-Heuristic methods

Classification of existing research using the Metaheuristic method

Classification of existing research using a Machine learning method

Classification of existing research using Statistical method

Details of different techniques, algorithms, workload data, approaches, and the researchers’ work are described below. In first section, the heuristic techniques and their different approaches are used to minimize energy usage. Many researchers work on the first fit decreasing (FFD), best fit decreasing (BFD), modified best fit decreasing, power-aware BFD, etc. Next, the metaheuristic techniques including swarm intelligence, evolutionary algorithm, nature-inspired algorithm, and physics-based algorithm are used to reduce energy consumption and to satisfy service level agreement. Machine learning techniques reinforcement learning, neural network (NN), support vector machine (SVM), and k-nearest neighbor (kNN) are elaborated in further section and finally statistical techniques using mean, standard deviation, regression, PPR gear, and ARIMA are explained.

Virtual machine management using heuristic techniques

A heuristic technique is a strategy for solving the problem, that is derived from the Greek term ‘eurisko,’ which means to search, find, or discover. It is about employing a practical technique that does not have to be perfect. Heuristic approaches reduce the time required to find a satisfactory answer. In cloud computing heuristic techniques are used for VM consolidation. In this approach, different researchers use FFD, BFD, MBFD, PABFD, and other algorithms for VM allocation, migration, and placement to reduce energy consumption.

Srikantaiah et al. 2008 [ 44 ], the virtual machine consolidation (VMC) problem was introduced as a bin packing problem. Researchers only examined two criteria: disc and CPU use. The analysis revealed that there is energy-performance compensation for consolidation, with the existence of optimal operating conditions. They constructed a cloud setting, collected data, and developed a bin packing issue using static random threshold values. Other resources, like memory and network, should also be measured, as they may be limiting resources for particular applications.

Beloglazov et al. (2010) [ 45 ] For VM consolidation with random data, researchers employed single threshold (ST), minimization of migration (MM), and bin packing strategies. The authors attempted to strike an ideal balance between energy savings and desired performance. They consolidated VMs based on current resource use, network topologies employed in VMs, and thermal status. An energy-aware resource scheduling system based on heuristics for VM allocation and live migration was suggested. The authors structured it as a bin packing issue and evaluated the effort using preset thresholds using the CloudSim toolbox. The results show that dynamic VM consolidation with adaptive thresholds outperforms static thresholds. Non-power-aware (NPA), dynamic voltage frequency scaling (DVFS), and ST methods were used to test the MM algorithm. Using energy savings, the MM algorithm outperformed ST, DVFS, and NPA by 23%, 66%, and 83%, respectively, with thresholds set at 30–70%, resulting in SLA breaches of 1.1%. The MM policy resulted in 6.7% SLA breaches and 43%, 74%, and 87% higher energy savings than the ST, DVFS, and NPA policies when the threshold value was kept at 50–90%.

Anton et al. (2011) [ 46 ] provided a heuristic approach for resource distribution that is energy efficient. The policy allocated resources to consumer apps in an energy-efficient manner while ensuring QoS by utilizing energy-efficient mapping heuristics using the consolidation of virtual machines. For VM placement, an improved form of the best fit decreasing modified BFD (MBFD) technique was utilized, as well as three double-threshold VM selection policies, random choice policy (RCP), highest potential growth, and MM. CPU usage data were produced at random by utilizing fixed criteria. The results in the CloudSim toolbox showed that energy consumption was reduced by 77% and 53%, respectively, as compared to NPA and DVFS policies, with SLA breaches of 5.4%.

Beloglazov et al. (2012) [ 43 ] Researchers proposed a dynamic VM consolidation approach because fixed thresholds are not feasible in a dynamic cloud environment. The authors reported dynamic threshold values by statistically assessing four histories of CPU use. The reallocation was carried out utilizing a dynamic threshold method. The MBFD technique was utilized to place the VMs. SLA-aware metrics were also examined. The results obtained by running the algorithm on the CloudSim toolkit with a genuine PlanetLab trace demonstrated the validity of the suggested framework. But in the model only single-core CPUs were used, and only a single-core resource CPU was tested.

Arani et al. (2018) [ 47 ] by providing a VM placement strategy, researchers concentrated on reducing energy use (VMP-BFD). VMs were mapped to hosts using an approach centred on the best fit decreasing approach, which significantly decreased energy use and SLA violations. The developed algorithm employed the theory of learning automata, correlation coefficients, and the ensemble forecast technique for VM allocation to hosts. The method assigned a VM to a host whose VMs had the least association with the chosen VM for placement. Compared to other reference policies, the results of the simulations on the CloudSim platform showed a big improvement in lowering the energy use and the SLA violations.

Wang et al. (2018) [ 48 ] focused on energy-efficient dynamic virtual machine consolidation (DVMC) by introducing an approach for virtual machine placement called “Space-Aware Best Fit Decreasing” (SABFD). The authors also created a VM selection strategy called “High CPU Utilization-based Migration VM Selection” (HS). The suggested system was evaluated in several ways by utilizing the CloudSim toolkit and the Planet Lab workload. The results showed that DVMC designs with a range of SABFD and HS produced the better results.

F.F. Moges et al. (2019) [ 27 ] proposed the OpenStack Neat framework’s VM placement method to address the issue of consolidation. They introduced VM placement methods that modify heuristics bin-packing to account for host energy efficiency. When linked to the reference algorithms PABFD and MBFD, the proposed algorithms improve energy proficiency. Depending on the host categories and workloads, the energy proficiency improvement over MBFD can be up to 67%. They also defined an innovative bin-packing method termed a “medium-fit” to avoid unnecessary SLAV and VM migrations. The MFPED (medium-fit power-efficient decreasing) offers a lower SLAV and VM migration rate compared to other VM placement methods. SLAV and VM relocations are reduced to 78% and 46%, respectively, when compared to MBFD, depending on the cloud scenario. They used CloudSim to test the suggested algorithms’ performance in three different data-center situations: heterogeneous, homogeneous, and default. Data workloads that execute in cloud centers are derived from PlanetLab and Bitbrains cloud traces.

Bhattacherjee et al. (2019) [ 49 ] for large historical data sets, proposed prediction technique that was accepted and employed in the current strategy known as the minimization of migration and dynamic thresholding system instead of static thresholds. The MBFD algorithm is used in prediction-based minimization of migration (PMM) to place the VMs. Markov chain learning is applied to formulate the past data for upcoming forecasting deployments. CloudSim 3.0.3 has been used to run rigorous simulations, and the outcomes show a decrease in cloud data center energy utilization.

Xialin Liu et al. (2020) [ 50 ] proposed dynamic consolidation by using migration thrashing. It prioritizes VMs with high dimensions and remarkably decreases migration thrashing. The degree of relocations required maintaining service-level agreements (SLAs) by keeping VMs prone to relocation thrashing on the identical physical servers rather than migrating. Their method improves the relocation thrashing measured around 28%, the number of movements measured around 21%, and the SLAV measured around 19%. When the server is overloaded, their solution detects VMs with sufficient capacity by restricting that VMs with excessive capacity are not transferred. Imitations of a wide-ranging research setting employing a workload data set from numerous PlanetLab VMs were used to validate the suggested techniques.

Saikishor Jangiti et al. (2020) [ 51 ] DRR-FFD and DRR-BinFill are cutting-edge VMC algorithms based on the concepts of FFD (first-fit decreasing) and DRR (dominant residual resource) that organize VMs based on a single VM resource. Researchers proposed an energy-efficient architecture — EMC2 — for an IaaS cloud service provider. The vector bin-packing techniques VMNeAR-E and VMNeAR-D are proposed. In a python context, simulation tests were conducted utilizing a dataset acquired from the EnergyStar® API for diverse physical servers. The suggested VMNeAR-D heuristic saved up to 3.318% of energy on the average across 40 schedules.

Garg et al. (2021) [ 52 ] provided load-aware three-gear THReshold (LATHR) and the MBFD algorithm to reduce overall energy consumption even though they improved service quality in terms of SLA. It produces promising results when used with a dynamic workload and a flexible count of virtual machines (1–290) on each host. The results of the projected work were evaluated concerning service level agreements (SLAs), energy utilization, the number of relocations against various numbers of virtual machines (VMs), and instruction energy ratio (IER). The proposed technique reduces SLA defilements (26%, 55%, and 39%) as well as energy consumption (12%, 17%, and 6%) when related to interquartile range (IQR), median absolute deviation (MAD), and double threshold overload acknowledgement strategies, respectively.

Alharbi et al. (2021) [ 53 ] improved existing research that manages data center resources using two independent layers: applications allotted to VMs and VM placement to hosts; both are bin packing problems. This sequential double-layered bin packing (Consec2LBP) solves issues easily and restricts added solution quality development. This research proposes an integrated ant colony optimization strategy to deal with the layers simultaneously to overcome this issue. It converts two-layer resource management into an optimization problem known as integrated double-layer bin packing (Int2LBP). Then, to solve this optimization challenge, a combined FFD technique known as Int2LBP_FFD is derived. To improve the quality of the result, a combined ant colony system, Int2LBP_ACS, has been developed, where the result of Int2LBP_FFD is used as a preliminary solution. In simulations of nine scales of data centers based on GTC data logs, integrated double-layer Int2LBP_FFD outperforms sequential Consec2LBP_FFD. They’ve also shown that Int2LBP_ACS is better than Int2LBP_FFD concerning energy investments. The Int2LBP_ACS and Int2LBP_FFD algorithms provide scalability.

T Kaur et al. (2022) [ 54 ] The Power Aware Energy Efficient Virtual Machine Migration (PAEEVMM) Method has been developed to migrate virtual machines in data centres depending on the temperature threshold value. Based on temperature, this approach moves the heavily loaded virtual machine to the less loaded virtual machine. The simulation was run on CloudSim Plus, and the outcomes are assessed against first fit algorithms. The experiment demonstrates that the suggested approach performs better in terms of CPU and electricity usage.

A brief description of the above detailed literature review and algorithms developed using heuristic methods with different workload data is given in Table 1 . Table  2 , summarises the work, method, and comparison with their benchmark methods/ algorithm to evaluate energy consumption. Figure  10 depicts the percentage difference in energy reduction or energy savings in graphical form. The implementation of these algorithms has been tested using different settings. The authors have already talked about the host specification, virtual machine description, datasets, simulators, and other criteria for comparing the proposed method to their benchmark algorithm.

Energy reduction using Heuristic techniques vs Benchmark

Virtual machine management using metaheuristic methods

A metaheuristic is a problem-solving strategy based on a heuristic method that is independent of the problem’s nature. A single-solution local search metaheuristic and a random search metaheuristic are the two types of metaheuristic methods. Metaheuristic approaches have been shown to produce near-optimal solutions in a reasonable amount of time and are problem-independent, allowing them to be used in a wide range of situations. It is advantageous in a cloud setting to locate a suboptimal solution quickly. Different metaheuristic techniques based on swarm intelligence, bio-inspired, physics-based, and evolutionary algorithms are used by researchers for VM consolidation to reduce energy consumption. This method was implemented for resource prediction, VM migration, VM placement, load balancing, etc.

Kousiouris et al. (2011) [ 55 ] worked on the analysis and performance of VM which depends on several parameters. They proposed the effects on VM performance prediction, persistent allocation proportions, VM co-placement, and instantaneous arrangement on the identical host. They applied a genetic algorithm (GA) to optimize an artificial neural network (ANN) and used linear regression to investigate degradation prediction.

Aryania et al. (2018) [ 56 ] proposed a technique using an ACS to resolve the VM consolidation (VMC) issue to reduce energy utilization in data centers. They took into account energy utilization through virtual machine migration. They presented an energy-aware VMC process based on an ACS to handle the VMC issue as a multi-objective optimization challenge. On the arbitrary workload in several circumstances, simulation findings showed that EVMC-ACS increased the number of sleeping hosts by 16% as related to ACS-VMC. Also, the suggested algorithm minimizes relocations by 89%, the power consumption during a migration by 91%, SLA violations by 79%, and overall energy consumption by 25% relative to ACS-VMC.

Goyal et al. (2019) [ 57 ] worked on PSO and CSA algorithms. The goal of optimizing energy utilization in the cloud is also addressed in the article. CloudSim simulators and common programming languages were utilized in their suggested work. Several performance measures, such as energy efficiency, response time, and execution time, were used to judge how well the work performs.

M Tarahomi et al. (2020) [ 58 ] approached micro-genetic method for choosing the right physical host for a virtual machine. Their simulations reveal that the micro-genetic method enhances power consumption relations. The suggested approach was tested using CloudSim and their result was related to the reference algorithms (genetic and PABFD VM provisioning algorithms) in various scenarios with the datasets of 10 working days. According to experimental results by the CloudSim framework, the micro-genetic system reduced power consumption.

Dubey et al. (2020) [ 59 ] suggested a virtual machine placement approach that reduces the makespan while reducing power consumption. The proposed technique was tested in the simulator CloudSim toolkit, and the findings proved that it exceeded typical work utilizing FCFS, Round-Robin, EERACC, and Random algorithms. The result shows that the recommended technique beats the other four mentioned methods regarding energy and power usage, server utilization, and makespan.

Barthwal et al. (2021) [ 60 ] proposed AntPu ACO meta-heuristic predicted utilization for dynamically placing VMs in the cloud data center to minimize SLAV and energy utilization (EU). In CloudSim, a simulated environment is created, and the PlanetLab dataset is chosen because of its real-world properties. The CPU usage of VMs in five-minute intervals is shown in this data set. To assess the results, extensive simulations were run, showing that the proposed approach offers a significant improvement in energy utilization and SLA compared with other methods. AntPu improves performance by satisfying SLA, QoS, EC, VM migration, and PM overloading constraints.

Mirmohseni et al. (2021) [ 61 ] combined the outcomes of the particle swarm genetic optimization (PSGO) process. The findings were improved and a viable solution for load balancing operations was introduced by combining the advantages of these two algorithms. Instead of arbitrarily assigning the beginning population or data set in the GA, the most acceptable outcome is obtained by giving the starting population in their proposed approach, load balancing PSGO Improve Resource Allocation (LBPSGORA). The LBPSGORA method is compared to GA, PSO, and a hybrid GA-PSO approach. This method outperformed similar methods in terms of execution cost, load balancing, and time to completion. With task changes, the hybrid GA-PSO approach performs similarly to the suggested method. The LBPSGORA technique is 7.32% more effective in makespan and 6.87% more effective in execution cost compared to the hybrid GA-PSO. LBPSGORA outperformed the hybrid GA-PSO by 8.42%, GA by 10.61%, and PSO by 11.71% in terms of load matching.

Alharbi et al. (2021) [ 53 ] improved existing research that manages data center resources using two independent layers: applications allotted to VMs and VM placement to hosts; both are bin packing problems. This sequential double-layered bin packing (Consec2LBP) makes easier the issue solving and restricts added solution quality development. This research proposes an integrated ant colony optimization strategy to deal with the layers simultaneously to overcome this issue. It converts two-layer resource management into an optimization problem known as integrated double-layer bin packing (Int2LBP). Then, to solve this optimization challenge, a combined FFD technique known as Int2LBP_FFD was developed. To improve the quality of the result, the combined ant colony system Int2LBP_ACS is refined further using the Int2LBP_FFD result as a preliminary solution. In simulations of data centers based on GTC data logs, Int2LBP_FFD outperforms Consec2LBP_FFD. They’ve also shown that Int2LBP_ACS is better than Int2LBP_FFD concerning energy investments. The Int2LBP_ACS and Int2LBP_FFD algorithms provide scalability.

Salami et al. (2021) [ 62 ] offer a virtual machine placement problem (VMPP) based on the cuckoo search (CS) algorithm. New cost and perturbation metrics have been created to increase the algorithm’s performance. Two well-known benchmark datasets were used to evaluate the suggested technique. The main objective is to organize virtual machines into actual machines to minimize the number of devices required. It beat the reordered grouping genetic algorithm and the FFD, BFD, and multiCSA, an older CS approach.

M. H. Sayadnavard et al. (2022) [ 63 ] approached a technique for dynamic VMC, which included a prediction model based on DTMC, a VM selection algorithm, and e-MOABC-based VM placement. Using this model in conjunction with the dependability model of PMs results in a more exact classification of PMs depending on their condition. Then, a multi-objective VM placement approach is proposed using the e-dominance-based multi-objective artificial bee colony algorithm to find the optimum VMs to PMs mapping, which can efficiently manage overall energy consumption, resource usage, and system performance to meet SLA and QoS requirements. By completing a performance assessment study with the CloudSim toolkit and PlanetLab workload traces, the proposed system is proved to be effective. The suggested technique greatly decreases energy usage while avoiding excessive VM migrations, according to a competitive analysis of the experimental findings. The investigation of various parameters reveals that the suggested approach outperforms other algorithms. MOABC-VMC decreases energy consumption by 11.35% and 35.25%, respectively, when compared to RE-VMC and LR-MMT.

S. Malik et al. (2022) [ 64 ], proposed Evolutionary Algorithms and Machine Learning Methods to Predict Resource Utilization in cloud data centers. The primary goal was to resolve the over-and under-provisioning problems. Over-provisioning of resources results in higher expenses and increased energy use. However, under-provisioning results in SLA violations and a decline in quality of service (QoS). The research focuses on functional link neural networks (FLNN) using hybrid Genetic Algorithms (GA) and Particle Swarm Optimization (PSO) for multi-resource usage prediction. The suggested model produces improved accuracy when compared to conventional procedures, according to experimental results using data from Google Cluster Traces. This study’s primary objective was to examine how well neural networks predicted multi-resource allocation. The proposed model predicts using FLNN and trains the network weights using a hybrid GA-PSO. To manage a large number of users, resources must be dynamically scaled for effective usage, low energy consumption, low cost, and higher quality of service (QoS).

A brief report of the above detailed literature review and algorithms mentioned using metaheuristic methods with different workload data is given in Table  3 . Table  4 , summarises researchers work, methods, and comparison with their benchmark algorithm to evaluate energy consumption. Figure  11 depicts the percentage difference in energy reduction or energy savings in graphical form. The implementation of these algorithms has been tested with different settings. About the host specification, virtual machine characteristics, workload datasets, simulators or tools, and other measures for comparing the proposed method to their benchmark algorithm has already been discussed earlier.

Energy reduction in Metaheuristic techniques vs Benchmark

Virtual machine management using machine learning techniques

Machine learning technique are approaches and set of technologies that use AI concepts. Machine learning enables researchers to use data to train a system on how to solve a problem using machine learning algorithms and improve over time. Machine learning is frequently classified by how an algorithm learns to improve its prediction accuracy. Supervised learning, unsupervised learning, semi-supervised learning, and reinforcement learning are the four fundamental methodologies. In a cloud computing environment reinforcement learning, neural network, k-nearest neighbor, and support vector machine algorithm are used by researchers to consume less amount of energy in cloud environment.

Jia et al. (2009) [ 65 ] have proposed a reinforcement learning method called VCONF, which automates the VM configuration process by addressing the system’s scalability and adaptability problems. By learning from repetitions with the environment, virtual machine configuration (VCONF) generates policies for the auto-configuration of VMs. This method achieves the best cloud setup while improving adaptability and scalability also. Experimental results demonstrated the system’s optimality in controlled problems, as well as its scalability and adaptability in a broader system. VCONF could be changed to a good configuration in seven steps and showed a 20% to 100% increase in throughput over simple RL approaches.

Vinh et al. (2010) [ 66 ] developed an energy-aware algorithm that uses a neural network (NN) to forecast upcoming load requirements built on previous data and reduces the number of hosts by shutting them down or restarting them as needed. Their research objective is to moderate the energy used in data centers. When the system load increases or decreases, the system turns on or off some hosts.

Niehorster et al. (2011) [ 67 ] have presented an approach for the provisioning of virtual machines using support vector machines (SVM). They created a self-configurable and self-optimized multi-agent system capable of learning its behaviour and estimating its cost. The system acquires performance models for various applications and develops a behaviour model, after which SVM is used to organize the data in the knowledge base.

Kousiouris et al. (2011) [ 55 ] depend on several parameters on VM performance prediction, persistent allocation proportions, VM co-placement, and instantaneous arrangement on the identical host. They used a genetic algorithm (GA) to improve an ANN and linear regression to study how well it could predict degradation.

Islam et al. (2012) [ 68 ] constructed a model for predicting future CPU resource requirements using the linear regression method. The input data set used historical data obtained by performing the Transaction Processing Performance Council (TPC), a typical client-server benchmark. To train the algorithm for prediction, the CPU utilization percentages of all VMs are used. They also used a neural network in the cloud for resource allocation and management. The neural network was trained with the back-propagation process, and experimental outcomes showed that NN-approximate predictions have a lower proportion error than LR-based predictions.

Cheng et al. (2012) [ 69 ] proposed a unified reinforcement learning technique for autonomously configuring virtual machines and their applications and adjusting the VM resources efficiently and providing quality service assurance. They came up with a good plan for running their research on Xen VMs using different workloads.

Farahnakian et al. (2013) [ 70 ] introduced a dynamic consolidation of virtual machines (DCVM) where the active number of hosts are minimized based on present and historical use. The k-nearest neighbour (KNN) method is used to forecast each host’s CPU utilization. To optimize dynamic VM consolidation, their prediction technique focuses on identifying overloading and underloading of hosts. The results indicated that their system consumes the least amount of energy while maintaining the SLA.

Farahnakian et al. (2014) [ 71 ] suggested a Reinforcement learning (RL) technique for dynamic consolidation of VM that uses a learning agent to find out the host’s power strategy. The agent selects the host to make it active or sleep. The RL learning agent optimizes the active host by learning system behavior. Experiments with PlanetLab workload traces show that their model lowers the cost of using energy, improves performance, and cuts down on SLA violations.

Minal et al. (2016) [ 72 ] Configure live VM migration using a support vector regression (SVR) model to forecast dirty pages using time series analysis. The service interruption time and migration duration were used to assess the performance of the live migration. They also created an ARIMA-based model, and findings show that SVR outperforms ARIMA in predicting dirty pages. Total pages transferred and migration time are the two most critical performance criteria for live migration in their proposed system.

Duggan et al. (2016) [ 73 ] developed a network-aware live migration technique that monitors bandwidth usage and takes appropriate action when there is network congestion based on experience Their structure functions as a decision support system, enabling a mediator to schedule VM migrations by determining the best time to do so. The amount of bandwidth available in the data center influences the migration process. According to their research findings, an agent in a cloud data center can learn available bandwidth during peak network capacity and schedule the migration of VMs from underutilized Hosts at the appropriate time using available bandwidth. They used the local regression approach to determine which hosts were overloaded. The Learning agent selects the best VM for migrating from an overloaded host while balancing migration and energy consumption. The findings of the research point to an autonomous VM selection method that can account for VM migration count and energy cost.

Duggan et al. (2017) [ 74 ] To create reliable predictions using time series data, researchers employed a recurrent neural network (RNN) to forecast future values of CPU consumption. They looked into the network’s accuracy for prediction with a deep effect. Experiments have shown that it is possible to get a very accurate estimate of CPU usage for dynamic data sets that change.

Qazi et al. (2017) [ 75 ] provided a real-time resource consumption prediction classification that takes actual resource usage and sends it to multiple buffers built on time and resource type. A system with real CPU utilization traces from a cloud data center with 120 servers used the autoregressive neural network method on data blocks where the data did not track a Gaussian distribution. The experimental findings suggest that AR-NN outperforms ARIMA for a given data set.

Shaw et al. (2017) [ 76 ] have presented the advanced RL consolidation agent method for VM allocation that is capable of optimizing VM circulation in the cloud data center while saving large amounts of energy and lowering SLA violations. They established a space for state-action. Action is defined as a combination of any host’s utilization rate and the size of the VM to be deployed, and state is well-defined as the entire active host as a percentage of the total host.

Sotiriadis et al. (2018) [ 77 ] proposed a VM scheduling strategy that uses extracted data from past VM and host resource utilizations to define host weights based on the resource utilization of hosted VMs on that host. They used SVM to classify VM states based on historical records. They used the resource utilization dataset (percentage of CPU, RAM, and disc usage) in the X-Y planes and expressed the data as vectors. The results of the experiments reveal that, through learning the system’s behavior, their method improved physical machine selection.

Mason et al. (2018) [ 78 ] using evolutionary NN, created a way to forecast the host’s CPU utilization. For network training, optimization approaches such as particle swarm optimization (PSO), differential evolution (DE), and covariance matrix adaptation evolutionary strategy (CMA-ES) are used. The outcomes of the experiments showed that CMA-ES performs better than other optimization strategies and trains networks to predict CPU consumption accurately.

Patel et al. (2019) [ 79 ] presented a load-balancing method based on energy-aware VM Migration. They perform it by assigning a lower and higher threshold to an individual host, which specifies whether the host is underloaded or overloaded. Before initiating the migrations, they used a prediction approach that predicts the demand on the host. Their process uses an artificial neural network (ANN) with the dynamic double threshold (DDThr) technique to predict VM movement and energy consumption while considering CPU utilization. Not only does it reduce the number of VM movements, but it also saves energy. Graphs comparing VM movement and energy utilization show that when ANN is combined with existing techniques, both VM movements, and energy utilization decrease slightly, saving a significant amount of electricity. To create a cloud environment, the CloudSim simulator was employed, and Matlab2015a was used to implement ANN. Based on the experiments, the proposed strategy uses less energy and has fewer migrations than the competitive approach.

Kumar et al. (2020) [ 80 ] provide a workload forecasting framework based on a NN (WFNN) model with supervised learning. To increase the predictive model’s learning efficiency, an upgraded and adaptable differential evolution method has been designed and developed. The algorithm determines the most appropriate crossover and mutation operators. Because of its adaptive nature in pattern learning from sampled data, the learning’s prediction accuracy and convergence rate have been seen to improve. The prediction model’s performance is assessed using real-world data traces from Google’s cluster and NASA’s Kennedy Space Center. A Python3 Jupyter notebook is used to implement the suggested model. The results are compared with other recent methods, and improvements of up to 97%, 91%, and 97.2% are observed over backpropagation, self-adaptive differential evolution, and average-based workload prediction techniques, respectively.

Saxena et al. (2021) [ 81 ] introduce an energy-efficient resource provisioning and management system to satisfy future applications’ dynamic demands. The proposed system addresses power consumption, performance, resource wastage, and QoS depletion by accurately matching the application’s expected resource demand with VM resource capacity. Consequently, condensing the whole load onto the smallest number of energy-efficient physical machines (PMs). The proposed work makes contributions in the form of online multi-resource feed-forward NN (OM-FNN) to predict resources, autoscaling of VMs, and allocation of scaled VMs on energy-efficient hosts. The suggested integrated solution has been rigorously evaluated using real resource usage traces from the Google cluster dataset, and it outperforms the other VMPs in terms of resource utilization and power savings by up to 21.12% and 88.5%, respectively. Also, the OM-FNN predictor is more accurate, takes less time, and uses less space than the single-input single-output feed-forward NN predictor.

Malik et al. (2022) [ 64 ] focuses on employing a hybrid Genetic Algorithm (GA) and Particle Swarm Optimization with a Functional Link Neural Network (FLNN) to anticipate the multi-resource utilization (CPU, memory, and network bandwidth). For resource usage prediction, the programme employs models from convolutional neural networks (CNN) and long short-term memory (LSTM). Experimental findings using Google cluster traces demonstrate that the suggested model outperforms conventional methods in terms of accuracy. This study’s major objective was to examine how well neural networks forecast the use of several resources. FLNN is used for prediction, while hybrid GA-PSO is used to train the network weights. Therefore, to manage a high number of users, the resources need to be scaled dynamically for optimal use, decreased energy consumption, and cost, with better quality of service (QoS).

A brief description of the above detailed literature review and algorithms developed using machine learning methods with different workload data is given in Table  5 . Table  6 , summarises the work, methods, and comparison with their benchmark algorithm to evaluate energy consumption by different researchers. Figure  12 depicts the percentage difference in energy reduction or energy savings in graphical form. The implementation of these algorithms has been tested for different settings. The authors have already mentioned the host specification, characteristics of virtual machine, workload datasets, simulators environment, and other criteria for comparing the proposed method to their benchmark algorithm.

Energy reduction in Machine Learning techniques vs Benchmark

Virtual machine management using statistical techniques

Statistical methods are used in research planning, analyzing, data collecting, meaningful interpretations, and reporting the findings of various virtual machine management. In cloud computing researchers work on mean, standard deviation, regression, ARIMA, PPRGear, etc. to detect overload and underload hosts, resource prediction, VM allocation, VM migration, and VM placement to save energy consumption.

Cao et al. (2012) [ 82 ] proposed strategies for dynamically combining VMs in a virtualized data center to reduce SLAV and energy utilization. The authors suggested detecting host overload, VM selection, and allocation strategy. The author’s uses mean and standard deviation CPU utilization metrics to determine overloaded hosts. The extension of the maximum correlation (MCE) strategy was utilized to select VMs for migration with mean and variance-related computations for VM allocation. Experiments using PlanetLab traces on CloudSim revealed that the new framework, which consists of the policies listed above, outperforms the previous policies in the requisites of energy utilization and overall QoS. However, it performed slightly worse in the requisites of energy utilization. As a result, managing the energy-performance trade-off is difficult.

Farahnakian et al. (2013) [ 83 ] Using PlanetLab historical data, a linear regression method was proposed to forecast the upcoming CPU use of the host (LIRCUP). Authors discovered a relationship between expected and current CPU use, where expected utilization is a dependent variable and current utilization is an autonomous variable. The LIRCUP algorithm detects overloaded hosts and maintains SLA and energy utilization by transferring some VMs from the overburdened hosts by comparing the expected CPU utilization value with the present utilization.

Nadjar et al. (2015) [ 84 ] present a decentralized scheduling strategy for DCVMs fitted with an auto-regressive integrated moving average (ARIMA) technique to progress resource provisioning by predicting VM resource usage to decrease SLAV and energy utilization in cloud data centers. Global Manager uses first fit decreasing, Cluster Manager uses max load VMP, and Local Manager uses the ARIMA model in their model. As a result, by utilizing ARIMA upper-bound prediction, it is possible to obtain a 90% reduction in migration and SLA violation rates and a 5.4% increase in energy savings. The CloudSim simulator was used to evaluate the method’s efficiency with recently proposed approaches that employed the same workload and experimental settings.

Ruan et al. (2015) [ 85 ] define performance-to-power ratio (PPR) as conscious virtual machine distribution in energy-efficient clouds. They describe “PPRGear,” a novel VM allocation mechanism that takes advantage of performance-to-power ratios for diverse types of hosts. PPRGear can ensure that the host devices use the least amount of power possible. Thus, this drastically lowers the energy usage with minimal performance loss. The proposed algorithm outperforms the competition.

Abdelsamea et al. (2017) [ 86 ] introduced multiple regression host overload detection (MRHOD) procedures that practice memory, CPU, and bandwidth to detect host overload and save energy significantly. They used a combination of factors to manage VMs while keeping energy consumption and SLAs low. They also created the hybrid local regression host overload detection (HLRHOD) method based on LR with hybrid variables. This algorithm outperforms single-factor methods.

Khoshkholghi et al. (2017) [ 87 ] by developing a method for overloaded host detection using iterative weighted linear regression (IWLR), which takes SLA constraints for data centers into consideration, researchers forecasted a dynamic, cost-effective, and energy-efficient management of virtual machines.

Hemavathy et al. (2019) [ 88 ] provide a prediction-based thermal aware server consolidation (PTASC) model, an integration technique that considers numeric and local architecture, as well as service level agreement. PTASC uses a statistical learning approach to consolidate servers (VM Migration). Cloud computing is a method of supplying essential resources by optimizing the usage of data-center resources, which raises energy costs. To reduce energy costs and enhance usage, new energy-efficient methods are proposed that reduce the overall energy consumption of computing and storage.

Lianpeng et al. (2019) [ 89 ] Based on the suggested robust simple linear regression (RobustSLR) prediction model, the authors developed a host overloading/underloading detection technique and a novel VM placement strategy for SLA-aware and energy-efficient virtual machine consolidation in cloud data centers. Unlike native linear regression, the proposed approaches update the forecast and slant toward over-prediction by including the error using eight ways of calculating the error. Researchers examined suggested techniques for the test by extending the CloudSim simulator with PlanetLab and random workload. The experimental findings demonstrate that the suggested approach can minimize SLA violation rates up to 99.16% and energy usage up to 25.43%.

Xialin Liu et al. (2020) [ 50 ] proposed dynamic consolidation using migration thrashing (MT), which prioritizes VMs with high dimensions, significantly decreasing MT. The degree of migrations required maintaining service level agreements (SLAs) by keeping VMs prone to relocation thrashing on the identical physical servers rather than migrating. Their method improves the relocation thrashing measured around 28%, the number of movements measured around 21%, and the SLAV measured around 19%. When the server is overloaded, their solution detects VMs with sufficient capacity by restricting the transfer of VMs with excessive capacity. The suggested techniques were proven to work by simulating large-scale research setting with a workload data set from many PlanetLab VMs.

Maryam C.-Samani et al. (2020) [ 90 ] suggested predictive consolidation of virtual machines (PCVM) using the ARIMA approach, which focuses on the DCVM over the fewest number of real servers. It also reduces the number of unnecessary migrations, detects PM overloading, and enforces SLAs using the ARIMA prediction model. Furthermore, the DVFS approach is utilized to determine the best frequency for heterogeneous physical devices. The experimental findings reveal that, the given framework greatly reduces energy usage while improving QoS characteristics as compared to various baseline techniques. The suggested solution was simulated using MATLAB and CloudSim with real-world PlanetLab workloads.

A brief description of the above detailed literature review and algorithms developed using statistical methods with different workload data is given in Table  7 . Table  8 , summarises the work, methods, and comparison with their benchmark algorithm to evaluate energy consumption by different researchers. Figure  13 depicts the percentage difference in energy reduction or energy savings in graphical form. The implementation of these algorithms has been tested with different settings. The authors have already discussed the specification of the host, characteristics of the virtual machine, workload datasets, simulator environment, and other criteria for comparing the proposed method to their benchmark algorithm.

Energy reduction in Statistical techniques vs Benchmark

Most of the above researchers have used PlanetLab workload traces, as shown in Table 9 , or Bitbrains workload traces, as shown in Table  10 for simulation in CloudSim, Matlab, Java, or other environments are given below. Half of the 800 physical nodes in PlanetLab’s simulated data center are HP ProLiant ML110G4 systems, while the other half are HP ProLiant ML110G5 systems, as depicted in Table  11 . For the smooth conduction of simulation, the power modeling has been configured in CloudSim as shown in Table  11 .

Result analysis

The result of the review paper work is to find the current research outcomes in energy-efficient resource management as stated in different sections. Table 2 , and Fig. 10  represent the saving of energy by up to 90% by different researchers using the heuristic method. The objectives addressed in the evaluation of this method were VM placement, VM allocation, VM migration, and resource utilization. In next section, the authors’ metaheuristic approaches were performed to address the objectives of VM consolidation, load balancing, resource management, PM overloading, VM migration, and VM placement. Metaheuristic methods in Table 4 and Fig.  11  showed an improvement in energy savings of up to 95%. Similarly, machine learning algorithms were presented to address the objectives of VM performance, prediction usage of resources, VM scheduling, dynamic consolidation, and resource management. With this approach, the reduction in energy consumption up to 88% has been shown as compared with other methods which has been illustrated in Table 6 and Fig. 12 . In the last approach mentioned in this paper, researchers used statistical methods to perform host overload/underload detection, dynamic consolidation of VMs, utilization prediction, and VM allocation. This approach reduces the energy consumption up to 84%, as shown in Table 8 and Fig. 13 . The outcomes of the review work are measured in terms of SLA, energy consumption, and the number of migrations against the different numbers of VMs. This review work focuses on energy utilization by different approaches in consolidating virtual machines. The results show that there has been an improvement in energy saving in the outcome of all the researchers by using different techniques. Other research outcomes include the use of integrated and combined approaches for utilization prediction, utilization, virtual machine consolidation, overload detection, VM selection, VM migration, and VM placement.

Major issues, suggestions, and future works

In this paper, the authors have outlined energy-efficient strategies for cloud computing. Several methods have been investigated, and their findings with parameters are listed in the tables. This paper can help people to find out the pros and cons of proposed energy-efficient algorithms that are motivated by researchers.

One of the main issues in cloud computing is using energy effectively, which necessitates the development of an eco-friendly environment. To meet SLAs, service providers must provide continuous power to data centers. This way, the data centers consume a large amount of energy and raise the cost of investment. However, the rising demand for cloud infrastructure has significantly increased the data center energy usage, which has become a crucial concern. As a result, energy-efficient solutions are necessary to reduce this energy utilization. Another significant challenge is the system’s reliability degradation because of the high frequency of consolidation and deploying VMs on PMs. Cloud efficiency is the capacity to make greater use of cloud resources at the lowest feasible cost. Other issues that must be addressed include scheduling challenges while PM-VM mapping for each user task, resource utilization prediction accuracy, overload, and underload host detection problems, and adaptive threshold estimation. Moreover, VM selection from the overloaded host, access to a real cloud data center to perform an experiment in a real environment, and improving user satisfaction along with the service providers are also various research challenges.

Most of the researchers have performed simulations in the CloudSim framework in an Infrastructure as a Service (IaaS) environment. In CloudSim, development tools, middleware technology, database management, resource computation, etc. help create and control cloud applications. Logical architecture is based on local and global managers. Cloud architecture is the organization of various components, including applications, databases, on-demand resources, storage, middleware, network devices, and software capabilities to provide services. Increased power use is a longstanding problem in today’s computer environment. The rise of applications using complex data has resulted in the construction of large data centers, which has increased the need for energy. According to the above analysis of energy-efficient strategies, the majority of the effort to minimize energy utilization in data centers is done by utilizing dynamic VM consolidation and resource management methods. Some researchers suggest multi-objective [ 91 ] algorithms that primarily address SLA, QoS, and resource usage while consuming less energy in cloud data centers. There has been little work done on heterogeneous physical devices, which requires considerable attention from the scientific community. Some major issues in current energy management techniques are prediction utilization of different resources [ 64 ]; mapping of VMs to PMs; host overload issues; VM selection from overloaded hosts; access to a real cloud data center; and VM placement. As VM placement is an NP-hard problem, metaheuristic approaches are the best suitable technique, which increases the complexity.

This research contributes significantly to provide important information related to the reduction of data center energy consumption, financial expenses, and the provision of QoS, hence assisting in the development of a strong, competitive cloud computing sector. This is especially crucial in the current green environment, where customers are becoming more environmentally concerned. Furthermore, according to recent research, data centers are a huge and rapidly rising energy-consumption sector of the economy, as well as a substantial source of CO 2 emissions. Also, the research done by [ 92 ] the use of blockchain technology and cloud solutions facilitates and improves not only the aggregation of data and secured access to it, but also has a huge impact on the reduction of CO2 emissions and reduces the carbon footprint. Hence, reducing greenhouse gas (GHG) emissions is an important energy policy objective for many nations, as well as achieving the United Nations Sustainable Development Goal (SDG) to transform the world by 2030. As a result, global research efforts should focus on the open problems described in this work to improve energy-efficient resource management approaches in cloud computing systems. Also, the researchers’ plan should be centred on reducing energy use and increasing resource use without hurting the performance of the system.

Summary and conclusion

Data centers consume a tremendous amount of electricity for computing user applications as well as cooling their equipment. Improving energy efficiency in data centers may reduce greenhouse gas (GHG) emissions, air pollution, and the amount of water utilized in power generation. So, minimizing energy consumption has been a key challenge in recent years. As a result, it is one of the key study areas in cloud computing. Many researchers are concentrating their efforts on lowering the energy usage of data center infrastructures. This review article looks at virtual and physical machine consolidation strategies using various methodologies to save energy. These strategies look at global energy conservation and resource management. As a result, resource usage increases, and data center energy consumption decreases. This paper aims to identify energy consumption research that has been conducted using various heuristics, metaheuristics, machine learning, and statistical methods. VM selection and migration, host CPU usage prediction, overload detection, and VM placement have been used to manage resources and efficient use energy. The energy savings achieved through various strategies are compared in this paper. Various researchers tested several strategies in cloud data centers to reduce energy consumption and SLAV. In the heuristic approach, researchers have saved from 5.4% to 90% of energy with their proposed method when compared with the existing methods. Similarly, the metaheuristic approach reduces energy consumption from 7.68% to 97%. The machine learning method and the statistical method save energy from 1.6% to 88.5%, and 5.4% to 84% respectively when compared to the benchmark approaches considering a variety of settings and parameters. So, energy saving can be maximized up to 90% using different approaches in respect of consolidation of VMs, prediction of workload traces, utilization of resources, host underload/overload detection, VM selection, VM migration, and VM placement. The results of this study could help researchers come up with new ideas for research that will add to their knowledge and make it easier to use energy efficiently in cloud computing. So, the overall outcome of this review paper is to understand different techniques of energy-saving applied in the cloud data centers. As the field of cloud computing is increasing day by day and its application area is increasing, focus must be on the different methods of energy consumption in cloud data centers.

Mell P, Grance T (2011) The NIST definition of cloud computing

Book   Google Scholar  

Wyld DC (2009) Moving to the cloud: an introduction to cloud computing in government, IBM center for the business of government

Google Scholar  

Sethi N (2019) The cloud environment and its basics: a review. Int J Comput Technol 6(1):82–88

Sotomayor B, Montero RS, Llorente IM, Foster I (2009) Virtual infrastructure management in private and hybrid clouds. IEEE Internet Comput 13(5):14–22

Article   Google Scholar  

Kaur T, Chana I (2018) GreenSched: an intelligent energy-aware scheduling for deadline-and-budget constrained cloud tasks, simulation modelling practice, and theory. 82:55–83. ISSN 1569-190X. https://doi.org/10.1016/j.simpat.2017.11.008

Albers S (2010) Energy-efficient algorithms. Commun ACM 53:86–96. https://doi.org/10.1145/1735223.1735245

Çağlar İ, Altılar DT (2022) Look-ahead energy-efficient VM allocation approach for data centers. J Cloud Comput 11(11). https://doi.org/10.1186/s13677-022-00281-x

Conti J, Holtberg P, Diefenderfer J, LaRose A, Turnure JT, Westfall L International Energy Outlook 2016 With projections to 2040, United States. https://doi.org/10.2172/1296780

Milani AS, Navimipour NJ (2016) Load balancing mechanisms and techniques in the cloud environments: systematic literature review and future trends. J Netw Comput Appl 71:86–98. ISSN 1084–8045. https://doi.org/10.1016/j.jnca.2016.06.003

Malik N, Sardaraz M, Tahir M, Shah B, Ali G, Moreira F (2021) Energy-efficient load balancing algorithm for workflow scheduling in cloud data centers using queuing and thresholds. Appl Sci 11(13):5849. https://doi.org/10.3390/app11135849

Singh S, Chana I (2016) Cloud resource provisioning: survey, status, and future research directions, knowledge and information system. 49:1005–1069. https://doi.org/10.1007/s10115-016-0922-3

Mohit Kumar SC, Sharma AG, Singh SP (2019) A comprehensive survey for scheduling techniques in cloud computing. J Netw Comput Appl 143:1–33. ISSN 1084–8045. https://doi.org/10.1016/j.jnca.2019.06.006

Singh S, Chana I (2016) A survey on resource scheduling in cloud computing: issues and challenges. J Grid Computing 14:217–264. https://doi.org/10.1007/s10723-015-9359-2

Chaurasia N, Kumar M, Chaudhry R et al (2021) Comprehensive survey on energy-aware server consolidation techniques in cloud computing. J Supercomput 77:11682–11737. https://doi.org/10.1007/s11227-021-03760-1

Bui DM, Tu NA, Huh EN (2021) Energy efficiency in cloud computing based on mixture power spectral density prediction. J Supercomput 77:2998–3023. https://doi.org/10.1007/s11227-020-03380-1

Bobroff N, Kochut A, Beaty K (2007) Dynamic placement of virtual machines for managing SLA violations, 10th IFIP/IEEE international symposium on integrated network management

Varrette S, Guzek M, Plugaru V, Besseron X, Bouvry P (2013) HPC performance and energy-efficiency of Xen, KVM and VMWare hypervisors, 25th international symposium on computer architecture and high-performance computing

Gelenbe E (2009) Steps toward self-aware networks. Commun ACM 52(7):66–75

Berl A, Gelenbe E, Girolama M, Giuliani G, Meer H, Dang MQ, Pentikousis K (2010) Energy-efficient cloud computing. Comput J 53(7):1045–1051

Luo, L., et al., (2012) A resource scheduling algorithm of cloud computing based on energy efficient optimization methods

Buyya R, Broberg J, Goscinski AM (2010) Cloud computing: principles and paradigms, Vol. 87. John Wiley & Sons

Zhang Q, Cheng L, Boutaba R (2010) Cloud computing: state-of-the-art and research challenges. J Internet Serv Appl 1:7–18. https://doi.org/10.1007/s13174-010-0007-6

Buyya R, Beloglazov A, Abawajy J (2010) Distributed, parallel, and cluster computing (cs. DC), energy-efficient management of data center resources for cloud computing: a vision, architectural elements, and open challenges. arXiv. https://doi.org/10.48550/arXiv.1006.0308

Buyya R, Ranjan R, Calheiros RN (2009) Modeling and simulation of scalable cloud computing environments and the CloudSim toolkit: challenges and opportunities, in 2009 international conference on high-performance computing & simulation

Calheiros RN et al (2011) CloudSim: a toolkit for modeling and simulation of cloud computing environments and evaluation of resource provisioning algorithms. Software Pract Exper 41(1):23–50

Article   MathSciNet   Google Scholar  

Park K, Pai VS (2006) CoMon: a mostly-scalable monitoring system for PlanetLab. ACM SIGOPS Oper Syst Rev 40(1):65–74

Moges FF, Abebe SL (2019) Energy-aware VM placement algorithms for the OpenStack neat consolidation framework. J Cloud Comput 8(1):1–14

Shen S, Van Beek V, Iosup A (2015) Statistical characterization of business-critical workloads hosted in cloud data centers. In: Proc - 2015 15th IEEE/ACM international symposium on cluster, cloud and grid computing 2015, pp 465–474. https://doi.org/10.1109/CCGrid.2015.60 arXiv:1302.5679v1

Chapter   Google Scholar  

Anoep S, Dumitrescu C, Epema D, Iosup A, Jan M, Li H, Wolters L. The grid workloads archive: bitbrains. http://gwa.ewi.tudelft.nl/datasets/gwa-t12-bitbrains

Amvrosiadis G, Park JW, Ganger GR, Gibson GA, Baseman E, DeBardeleben N (2018) On the diversity of cluster workloads and its impact on research results. In: Proceedings of USENIX ATC

Reiss C, Tumanov A, Ganger GR, Katz RH, Kozuch MA (2012) Heterogeneity and dynamicity of clouds at scale: google trace analysis. In: Proceedings of ACM SoCC, pp 1–13

Liu Q, Zhibin Y (2018) The elasticity and plasticity in semicontainerized co-locating cloud workload: a view from Alibaba trace. In: Proceedings of ACM SoCC, pp 347–360

Shahrad M, Fonseca R, Goiri Í, Chaudhry G, Batum P, Cooke J, Laureano E, Tresness C, Russinovich M, Bianchini R (2020) Serverless in the wild: characterizing and optimizing the serverless workload at a large cloud provider. In: Proceedings of USENIX ATC, pp 205–218

Luo S, Xu H, Lu C, Ye K, Xu G, Zhang L, Yu D, He J, Xu C (2021) Characterizing microservice dependency and performance: Alibaba trace analysis, SoCC'21

Buyya R, Beloglazov A, Abawajy J (2010) Energy-efficient management of data center resources for cloud computing: a vision, architectural elements, and open challenges

Katal A, Dahiya S, Choudhury T (2022) Energy efficiency in cloud computing data centers: a survey on software technologies. Clust Comput. https://doi.org/10.1007/s10586-022-03713-0

Zhang Y, Liu J (2022) Prediction of overall energy consumption of data centers in different locations, Sensors. 22:3704. https://doi.org/10.3390/s22103704

Teng F et al (2017) Energy efficiency of VM consolidation in IaaS clouds. J Supercomput 73(2):782–809

Zhou Z et al (2018) Minimizing SLA violation and power consumption in cloud data centers using adaptive energy-aware algorithms. Future Gene Comput Syst 86:836–850

Khosravi A (2017) Energy, and carbon-efficient resource management in geographically distributed cloud data centers

Kaur T, Chana I (2015) Energy efficiency techniques in cloud computing: a survey and taxonomy. ACM Comput Surv 48(2):1–46

Clark C, Fraser K, SH, J., & Hansen, E. (2005) Warfield, live migration of virtual machines. In: Jul C, Limpach I, Pratt A (eds) In proceedings of the 2nd ACM/USENIX symposium on networked systems design and implementation (NSDI), pp 273–286

Beloglazov A, Buyya R (2012) Optimal online deterministic algorithms and adaptive heuristics for energy and performance efficient dynamic consolidation of virtual machines in cloud data centers, concurrency and computation. Pract Exper 24(13):1397–1420

Srikantaiah S, Kansal A, Zhao F (2008) Energy-aware consolidation for cloud computing. Clust Comput 12:1–15

Beloglazov A, Buyya R (2010) Energy-efficient resource management in virtualized cloud data centers,10th IEEE/ACM international conference on cluster, cloud and grid computing

Beloglazov A, Abawajy J, Buyya R (2012) Energy-aware resource allocation heuristics for efficient management of data centers for cloud computing. Future Gene Comput Syst 28(5):755–768

Ghobaei-Arani M et al (2018) A learning-based approach for virtual machine placement in cloud data centers. Int J Commun Syst 31(8):1–18

Wang H, Tianfield H (2018) Energy-aware dynamic virtual machine consolidation for cloud datacenters. IEEE Access 6:15259–15273

Bhattacherjee S et al (2020) Energy-efficient migration techniques for cloud environment: a step toward green computing. J Supercomput 76(7):5192–5220

Liu X et al (2020) Virtual machine consolidation with minimization of migration thrashing for cloud data centers. Math Probl Eng 2020:1–13

Jangiti S, Shankar Sriram VS (2020) EMC2: energy-efficient and multi-resource- fairness virtual machine consolidation in cloud data centers. Sustain Comput: Inform Syst 27:100414. ISSN 2210–5379. https://doi.org/10.1016/j.suscom.2020.100414

Garg V, Jindal B (2021) Energy-efficient virtual machine migration approach with SLA conservation in cloud computing. J Cent South Univ 28(3):760–770

Alharbi F et al (2021) Simultaneous application assignment and virtual machine placement via ant colony optimization for energy-efficient enterprise data centers. Clust Comput 24:1255–1275

Kaur T, Kumar A (2022) Power aware energy efficient virtual machine migration (PAEEVMM) in cloud computing. In: Satyanarayana C, Samanta D, Gao XZ, Kapoor RK (eds) High performance computing and networking. Springer, Singapore. https://doi.org/10.1007/978-981-16-9885-9_46 Lecture notes in electrical engineering, vol 853

Kousiouris G, Cucinotta T, Varvarigou T (2011) The effects of scheduling, workload type, and consolidation scenarios on virtual machine performance and their prediction through optimized artificial neural networks. J Syst Softw 84(8):1270–1291

Aryania A, Aghdasi HS, Khanli LM (2018) Energy-aware virtual machine consolidation algorithm based on ant colony system. J Grid Comput 16(3):477–491

Goyal et al (2019) An optimized model for energy efficiency on cloud system using PSO & CUCKOO search algorithm. Int J Innov Technol Explor Eng 8(9S):2278–3075

Tarahomi M, Izadi M, Ghobaei-Arani M (2020) An efficient power-aware VM allocation mechanism in cloud data centers: a micro genetic-based approach. Clust Comput 24(2):919–934

Dubey et al., (2020), A simulated annealing based energy-efficient vm placement policy in cloud computing, 2020 international conference on emerging trends in information technology and engineering (ic-ETITE)

Barthwal V, Rauthan MMS (2021) AntPu: a meta-heuristic approach for energy-efficient and SLA aware management of virtual machines in cloud computing. Memetic Computing 13(1):91–110

Mirmohseni SM, Javadpour A, Tang C (2021) LBPSGORA: create load balancing with particle swarm genetic optimization algorithm to improve resource allocation and energy consumption in clouds networks. Math Probl Eng 2021:1–15.

Salami et al., (2021). An energy-efficient cuckoo search algorithm for virtual machine placement in cloud computing data centers. The Journal of Supercomputing 77(11):13330–13357.

Sayadnavard MH, Haghighat AT, Rahmani AM (2022) A multi-objective approach for energy-efficient and reliable dynamic VM consolidation in cloud data centers, engineering science and technology. Int J 26:100995. ISSN 2215–0986. https://doi.org/10.1016/j.jestch.2021.04.014

Malik S, Tahir M, Sardaraz M, Alourani A (2022) A resource utilization prediction model for cloud data centers using evolutionary algorithms and machine learning techniques. Appl Sci 12(4):2160. https://doi.org/10.3390/app12042160

Rao J et al (2009) Vconf: a reinforcement learning approach to virtual machines auto-configuration. In: Proceedings of the 6th international conference on Autonomic computing

Duy TVT, Sato Y, Inoguchi Y (2010) Performance evaluation of a green scheduling algorithm for energy savings in cloud computing, IEEE international symposium on parallel & distributed processing, workshops and Ph.D. forum (IPDPSW)

Niehorster, O., et al., (2011), Autonomic resource management with support vector machines, IEEE/ACM 12th international conference on grid computing

Islam S et al (2012) Empirical prediction models for adaptive resource provisioning in the cloud. Future Gene Comput Syst 28(1):155–162

Xu C-Z, Rao J, Bu X (2012) URL: a unified reinforcement learning approach for autonomic cloud management. J Parallel Distrib Comput 72(2):95–105

Farahnakian F et al (2013) Energy-aware consolidation algorithm based on k-nearest neighbor regression for cloud data centers, Department of IT, University of Turku. IEEE/ACM 6th International Conference on Utility and Cloud Computing, Finland

Farahnakian F, Liljeberg P, Plosila J (2014) Energy-efficient virtual machines consolidation in cloud data centers using reinforcement learning. In: 22nd Euromicro International Conference on Parallel, Distributed, and Network-Based Processing, pp 500–507

Patel M, Chaudhary S, Garg S (2016) Machine learning-based statistical prediction model for improving the performance of live virtual machine migration. J Eng 2016:1–9

Duggan M et al (2017) A reinforcement learning approach for the scheduling of live migration from underutilized hosts. Memetic Computing 9(4):283–293

Duggan M et al (2017) Predicting host CPU utilization in cloud computing using recurrent neural networks. In: In 2017 12th international conference for internet technology and secured transactions (ICITST)

Zia Ullah Q, Hassan S, Khan GM (2017) Adaptive resource utilization prediction system for infrastructure as a service cloud. Comput Intell Neurosci 2017:1–12

Shaw R, Howley E, Barrett E (2017) An advanced reinforcement learning approach for energy-aware virtual machine consolidation in cloud data centers. In: 12th international conference for internet technology and secured transactions (ICITST)

Sotiriadis S, Bessis N, Buyya R (2018) Self-managed virtual machine scheduling in cloud systems. Inf Sci 433:381–400

Mason K et al (2018) Predicting host CPU utilization in the cloud using evolutionary neural networks. Future Gene Comput Syst 86:162–173

Patel D, Gupta RK, Pateriya R (2019) Energy-aware prediction-based load balancing approach with VM migration for the cloud environment. In: Data, engineering and applications. Springer, pp 59–74

Kumar J, Saxena D, Singh AK, Mohan A (2020) Biphase adaptive learning-based neural network model for cloud datacenter workload forecasting. Soft Comput 24(19):14593–14610. https://doi.org/10.1007/s00500-020-04808-9

Saxena D, Singh AK (2021) A proactive autoscaling and energy-efficient VM allocation framework using online multi-resource neural network for cloud data center. Neurocomputing 426:248–264. ISSN 0925–2312. https://doi.org/10.1016/j.neucom.2020.08.076

Cao Z, Dong S (2012) Dynamic VM consolidation for energy-aware and SLA violation reduction in cloud computing. In: IEEE 13th international conference on parallel and distributed computing, applications and Technologies

Farahnakian F, Liljeberg P, Plosila J (2013) LiRCUP: linear regression-based CPU usage prediction algorithm for live migration of virtual machines in data centers. In: 39th Euromicro conference on software engineering and advanced applications (SEAA), pp 358–364

Nadjar A, Abrishami S, Deldari H (2015) Hierarchical VM scheduling to improve energy and performance efficiency in IaaS Cloud data centers. In: 5th international conference on computer and knowledge engineering (iccke)

Ruan X, Chen H (2015) Performance-to-power ratio aware virtual machine (VM) allocation in energy-efficient clouds. In: IEEE international conference on cluster computing, pp 264–273

Abdelsamea A et al (2017) Virtual machine consolidation enhancement using hybrid regression algorithms. Egypt Inform J 18(3):161–170

Khoshkholghi MA et al (2017) Energy-efficient algorithms for dynamic virtual machine consolidation in cloud data centers. IEEE Access 5:10709–10722

Hemavathy M, Anitha R (2019) Green aware based VM-placement in cloud computing environment using extended multiple linear regression model. In: Emerging trends in computing and expert technology, COMET 2019. Springer, Cham. https://doi.org/10.1007/978-3-030-32150-5_53 Lecture notes on data engineering and communications technologies, vol 35

Li L, Dong J, Zuo D, Wu J (2019) SLA-aware and energy-efficient VM consolidation in cloud data centers using robust linear regression prediction model. IEEE Access 7:9490–9500

Chehelgerdi-Samani M, Safi-Esfahani F (2020) PCVM.ARIMA: predictive consolidation of virtual machines applying the ARIMA method. J Supercomput 77:2172–2206. https://doi.org/10.1007/s11227-020-03354-3

Chen J, Du T, Xiao G (2021) A multi-objective optimization for resource allocation of emergent demands in cloud computing. J Cloud Computing 10(20):1–17

Karaszewski R, Modrzyński P, Modrzyńska J (2021) The use of Blockchain Technology in Public Sector Entities Management: an example of security and energy efficiency in cloud computing data processing. Energies 14(7):1873. https://doi.org/10.3390/en14071873

Download references

Acknowledgments

The authors would like to thank the editor and reviewers for their suggestions and improvement to the manuscript.

Not Applicable.

The authors are not affiliated with any entity that has a direct or indirect financial interest in the subject matter covered in the paper.

Author information

Authors and affiliations.

Hemvati Nandan Bahuguna Garhwal University, Srinagar Garhwal, Uttarakhand, India

Suraj Singh Panwar, M. M. S. Rauthan & Varun Barthwal

You can also search for this author in PubMed   Google Scholar

Contributions

Throughout all of these stages of the study, the writers contributed equally. The final manuscript was reviewed and approved by all authors. Suraj Singh Panwar and MMS Rauthan contributed to the idea and design, data analysis, and interpretation; Suraj Singh Panwar and Varun Barthwal authored the paper, conducted a literature review, and critically edited it for key intellectual content, and confirmed the final version.

Corresponding author

Correspondence to Suraj Singh Panwar .

Ethics declarations

Competing interests.

This manuscript has not been submitted to or is currently being reviewed by any other journal or publication platform.

Additional information

Publisher’s note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ .

Reprints and permissions

About this article

Cite this article.

Panwar, S.S., Rauthan, M.M.S. & Barthwal, V. A systematic review on effective energy utilization management strategies in cloud data centers. J Cloud Comp 11 , 95 (2022). https://doi.org/10.1186/s13677-022-00368-5

Download citation

Received : 03 March 2022

Accepted : 12 November 2022

Published : 17 December 2022

DOI : https://doi.org/10.1186/s13677-022-00368-5

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Cloud computing
• Data center
• Virtual machine
• Energy consumption

Information

• Author Services

Initiatives

You are accessing a machine-readable page. In order to be human-readable, please install an RSS reader.

All articles published by MDPI are made immediately available worldwide under an open access license. No special permission is required to reuse all or part of the article published by MDPI, including figures and tables. For articles published under an open access Creative Common CC BY license, any part of the article may be reused without permission provided that the original article is clearly cited. For more information, please refer to https://www.mdpi.com/openaccess .

Feature papers represent the most advanced research with significant potential for high impact in the field. A Feature Paper should be a substantial original Article that involves several techniques or approaches, provides an outlook for future research directions and describes possible research applications.

Feature papers are submitted upon individual invitation or recommendation by the scientific editors and must receive positive feedback from the reviewers.

Editor’s Choice articles are based on recommendations by the scientific editors of MDPI journals from around the world. Editors select a small number of articles recently published in the journal that they believe will be particularly interesting to readers, or important in the respective research area. The aim is to provide a snapshot of some of the most exciting work published in the various research areas of the journal.

Original Submission Date Received: .

• Active Journals
• Find a Journal
• Proceedings Series
• For Authors
• For Reviewers
• For Editors
• For Librarians
• For Publishers
• For Societies
• For Conference Organizers
• Open Access Policy
• Institutional Open Access Program
• Special Issues Guidelines
• Editorial Process
• Research and Publication Ethics
• Article Processing Charges
• Testimonials
• Preprints.org
• SciProfiles
• Encyclopedia

Article Menu

• Subscribe SciFeed
• Recommended Articles
• PubMed/Medline
• Google Scholar
• on Google Scholar
• Table of Contents

Find support for a specific problem in the support section of our website.

Please let us know what you think of our products and services.

Visit our dedicated information section to learn more about MDPI.

JSmol Viewer

Prediction of overall energy consumption of data centers in different locations.

1. Introduction

2. related work, 3. energy consumption prediction framework, 3.1. pue prediction, 3.2. it equipment energy consumption model, 3.3. calculation of the total energy consumption and related analysis, 4. evaluation, 4.1. sensitivity results, 4.2. annual pue estimation analysis, 4.3. carbon emissions prediction, 4.4. electricity cost estimation, 5. conclusions, author contributions, data availability statement, conflicts of interest.

Click here to enlarge figure

• Miller, R. The Sustainability Imperative: Green Data Centers and Our Cloudy Future. 2021. Available online: https://datacenterfrontier.com/green-data-center-imperative (accessed on 25 March 2022).
• Liu, Y.; Wei, X.; Xiao, J.; Liu, Z.; Xu, Y.; Tian, Y. Energy consumption and emission mitigation prediction based on data center traffic and PUE for global data centers. Glob. Energy Interconnect. 2020 , 3 , 272–282. [ Google Scholar ] [ CrossRef ]
• Ahmed, K.M.U.; Bollen, M.H.; Alvarez, M. A Review of Data Centers Energy Consumption And Reliability Modeling. IEEE Access 2021 , 9 , 152536–152563. [ Google Scholar ] [ CrossRef ]
• Dayarathna, M.; Wen, Y.; Fan, R. Data Center Energy Consumption Modeling: A survey. IEEE Commun. Surv. Tutor. 2016 , 18 , 732–794. [ Google Scholar ] [ CrossRef ]
• Rambo, J.; Joshi, Y. Modeling of data center airflow and heat transfer: State of the art and future trends. Distrib. Parallel Databases 2007 , 21 , 193–225. [ Google Scholar ] [ CrossRef ]
• Beloglazov, A.; Buyya, R.; Lee, Y.C.; Zomaya, A. A taxonomy and survey of energy-efficient data centers and cloud computing systems. In Advances in Computers ; Elsevier: Amsterdam, The Netherlands, 2011; Volume 82, pp. 47–111. [ Google Scholar ]
• Jiye, W.; Biyu, Z.; Fa, Z.; Xiang, S.; Nan, Z.; Zhiyong, L. Data center energy consumption models and energy efficient algorithms. J. Comput. Res. Dev. 2019 , 56 , 1587. [ Google Scholar ]
• Niederreiter, H.; Winterhof, A. Quasi-monte carlo methods. In Applied Number Theory ; Springer: Berlin/Heidelberg, Germany, 2015; pp. 185–306. [ Google Scholar ]
• Nossent, J.; Elsen, P.; Bauwens, W. Sobol’sensitivity analysis of a complex environmental model. Environ. Model. Softw. 2011 , 26 , 1515–1525. [ Google Scholar ] [ CrossRef ]
• Zhang, H.; Shao, S.; Xu, H.; Zou, H.; Tian, C. Free cooling of data centers: A review. Renew. Sustain. Energy Rev. 2014 , 35 , 171–182. [ Google Scholar ] [ CrossRef ]
• Yuventi, J.; Mehdizadeh, R. A critical analysis of Power Usage Effectiveness and its use in communicating data center energy consumption. Energy Build. 2013 , 64 , 90–94. [ Google Scholar ] [ CrossRef ]
• Choo, K.; Galante, R.M.; Ohadi, M.M. Energy consumption analysis of a medium-size primary data center in an academic campus. Energy Build. 2014 , 76 , 414–421. [ Google Scholar ] [ CrossRef ]
• Roy, S.; Rudra, A.; Verma, A. An energy complexity model for algorithms. In Proceedings of the 4th conference on Innovations in Theoretical Computer Science, Berkeley, CA, USA, 9–12 January 2013; pp. 283–304. [ Google Scholar ]
• Tudor, B.M.; Teo, Y.M. On understanding the energy consumption of arm-based multicore servers. In Proceedings of the ACM Sigmetrics/International Conference on Measurement and Modeling of Computer Systems, London, UK, 11–15 June 2013; pp. 267–278. [ Google Scholar ]
• Ge, R.; Feng, X.; Cameron, K.W. Modeling and evaluating energy-performance efficiency of parallel processing on multicore based power aware systems. In Proceedings of the 2009 IEEE International Symposium on Parallel & Distributed Processing, Chengdu, China, 10–12 August 2009; pp. 1–8. [ Google Scholar ]
• Song, S.L.; Barker, K.; Kerbyson, D. Unified performance and power modeling of scientific workloads. In Proceedings of the 1st International Workshop on Energy Efficient Supercomputing, Denver, CO, USA, 17–21 November 2013; pp. 1–8. [ Google Scholar ]
• Gao, Y.; Guan, H.; Qi, Z.; Wang, B.; Liu, L. Quality of service aware power management for virtualized data centers. J. Syst. Archit. 2013 , 59 , 245–259. [ Google Scholar ] [ CrossRef ]
• Fan, X.; Weber, W.D.; Barroso, L.A. Power provisioning for a warehouse-sized computer. ACM Sigarch Comput. Archit. News 2007 , 35 , 13–23. [ Google Scholar ] [ CrossRef ]
• Gu, C.; Li, Z.; Huang, H.; Jia, X. Energy efficient scheduling of servers with multi-sleep modes for cloud data center. IEEE Trans. Cloud Comput. 2018 , 8 , 833–846. [ Google Scholar ] [ CrossRef ]
• Kim, W.; Gupta, M.S.; Wei, G.Y.; Brooks, D. System level analysis of fast, per-core DVFS using on-chip switching regulators. In Proceedings of the 2008 IEEE 14th International Symposium on High Performance Computer Architecture, Salt Lake City, UT, USA, 16–20 February 2008; pp. 123–134. [ Google Scholar ]
• Van Heddeghem, W.; Lambert, S.; Lannoo, B.; Colle, D.; Pickavet, M.; Demeester, P. Trends in worldwide ICT electricity consumption from 2007 to 2012. Comput. Commun. 2014 , 50 , 64–76. [ Google Scholar ] [ CrossRef ] [ Green Version ]
• Chi, C.; Ji, K.; Marahatta, A.; Song, P.; Zhang, F.; Liu, Z. Jointly optimizing the IT and cooling systems for data center energy efficiency based on multi-agent deep reinforcement learning. In Proceedings of the Eleventh ACM International Conference on Future Energy Systems, Melbourne, QC, Australia, 22–26 June 2020; pp. 489–495. [ Google Scholar ]
• Wang, J.; Zhou, B.; Liu, W.; Hu, S. Research progress and development trend of cross-layer energy efficiency optimization in data centers. Sci. Sin. Inf. 2020 , 50 , 1–24. [ Google Scholar ] [ CrossRef ] [ Green Version ]
• Wan, J.; Gui, X.; Kasahara, S.; Zhang, Y.; Zhang, R. Air flow measurement and management for improving cooling and energy efficiency in raised-floor data centers: A survey. IEEE Access 2018 , 6 , 48867–48901. [ Google Scholar ] [ CrossRef ]
• Li, L.; Liang, C.J.M.; Liu, J.; Nath, S.; Terzis, A.; Faloutsos, C. Thermocast: A cyber-physical forecasting model for datacenters. In Proceedings of the 17th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Diego, CA, USA, 21–24 August 2011; pp. 1370–1378. [ Google Scholar ]
• Hsu, Y.F.; Matsuda, K.; Matsuoka, M. Self-aware workload forecasting in data center power prediction. In Proceedings of the 2018 18th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing (CCGrid), Washington, DC, USA, 1–4 May 2018; pp. 321–330. [ Google Scholar ]
• Gao, J. Machine Learning Applications for Data Center Optimization. 2014. Available online: https://storage.googleapis.com/pub-tools-public-publication-data/pdf/42542.pdf (accessed on 11 March 2022).
• Brady, G.A.; Kapur, N.; Summers, J.L.; Thompson, H.M. A case study and critical assessment in calculating power usage effectiveness for a data centre. Energy Convers. Manag. 2013 , 76 , 155–161. [ Google Scholar ] [ CrossRef ]
• Gozcu, O.; Ozada, B.; Carfi, M.U.; Erden, H.S. Worldwide energy analysis of major free cooling methods for data centers. In Proceedings of the 2017 16th IEEE Intersociety Conference on Thermal and Thermomechanical Phenomena in Electronic Systems (ITherm), Marina, San Diego, CA, USA, 31 May–1 June 2017; pp. 968–976. [ Google Scholar ]
• Chi, Y.Q.; Summers, J.; Hopton, P.; Deakin, K.; Real, A.; Kapur, N.; Thompson, H. Case study of a data centre using enclosed, immersed, direct liquid-cooled servers. In Proceedings of the 2014 Semiconductor Thermal Measurement and Management Symposium (SEMI-THERM), San Jose, CA, USA, 9–13 March 2014; pp. 164–173. [ Google Scholar ]
• Sharafi, M.; ElMekkawy, T.Y.; Bibeau, E.L. Optimal design of hybrid renewable energy systems in buildings with low to high renewable energy ratio. Renew. Energy 2015 , 83 , 1026–1042. [ Google Scholar ] [ CrossRef ]
• Shuja, J.; Gani, A.; Shamshirband, S.; Ahmad, R.W.; Bilal, K. Sustainable cloud data centers: A survey of enabling techniques and technologies. Renew. Sustain. Energy Rev. 2016 , 62 , 195–214. [ Google Scholar ] [ CrossRef ]
• Ahmad, R.W.; Gani, A.; Hamid, S.H.A.; Shiraz, M.; Yousafzai, A.; Xia, F. A survey on virtual machine migration and server consolidation frameworks for cloud data centers. J. Netw. Comput. Appl. 2015 , 52 , 11–25. [ Google Scholar ] [ CrossRef ]
• Nehra, P.; Nagaraju, A. Sustainable Energy Consumption Modeling for Cloud Data Centers. In Proceedings of the 2019 IEEE 5th International Conference for Convergence in Technology (I2CT), Bombay, India, 29–31 March 2019; pp. 1–4. [ Google Scholar ]
• Yamini, B.; Selvi, D.V. Cloud virtualization: A potential way to reduce global warming. In Proceedings of the Recent Advances in Space Technology Services and Climate Change 2010 (RSTS & CC-2010), Chennai, India, 13–15 November 2010; pp. 55–57. [ Google Scholar ]
• Zhang, Z.; Fu, S. Characterizing power and energy usage in cloud computing systems. In Proceedings of the 2011 IEEE Third International Conference on Cloud Computing Technology and Science, Athens, Greece, 29 November–1 December 2011; pp. 146–153. [ Google Scholar ]
• Lei, N.; Masanet, E. Statistical analysis for predicting location-specific data center PUE and its improvement potential. Energy 2020 , 201 , 117556. [ Google Scholar ] [ CrossRef ]
• Luo, L.; Wu, W.J.; Zhang, F. Energy modeling based on cloud data center. J. Softw. 2014 , 25 , 1371–1387. [ Google Scholar ]
• Shahid, A.; Fahad, M.; Manumachu, R.R.; Lastovetsky, A. A comparative study of techniques for energy predictive modeling using performance monitoring counters on modern multicore CPUs. IEEE Access 2020 , 8 , 143306–143332. [ Google Scholar ] [ CrossRef ]
• Shahid, A.; Fahad, M.; Manumachu, R.R.; Lastovetsky, A. Energy Predictive Models of Computing: Theory, Practical Implications and Experimental Analysis on Multicore Processors. IEEE Access 2021 , 9 , 63149–63172. [ Google Scholar ] [ CrossRef ]
• Azevedo, D.; Patterson, M.; Pouchet, J.; Tipley, R. Carbon usage effectiveness (CUE): A green grid data center sustainability metric. Green Grid 2010 , 32 , 4–5. [ Google Scholar ]
• Deng, W.; Liu, F.M.; Jin, H.; Li, D. Leveraging renewable energy in cloud computing datacenters: State of the art and future research. Jisuanji Xuebao (Chin. J. Comput.) 2013 , 36 , 582–598. [ Google Scholar ] [ CrossRef ]
• Efron, B.; Tibshirani, R.J. An Introduction to the Bootstrap ; CRC Press: Boca Raton, FL, USA, 1994. [ Google Scholar ]
• Sobol, I.M. Global sensitivity indices for nonlinear mathematical models and their Monte Carlo estimates. Math. Comput. Simul. 2001 , 55 , 271–280. [ Google Scholar ] [ CrossRef ]
• GEIDCO. China Carbon Neutrality Research Report to 2060. 2021. Available online: http://www.geidco.org.cn/html/qqnyhlw/zt20210120_1/index.html (accessed on 1 April 2022).

Share and Cite

Zhang, Y.; Liu, J. Prediction of Overall Energy Consumption of Data Centers in Different Locations. Sensors 2022 , 22 , 3704. https://doi.org/10.3390/s22103704

Zhang Y, Liu J. Prediction of Overall Energy Consumption of Data Centers in Different Locations. Sensors . 2022; 22(10):3704. https://doi.org/10.3390/s22103704

Zhang, Yiliu, and Jie Liu. 2022. "Prediction of Overall Energy Consumption of Data Centers in Different Locations" Sensors 22, no. 10: 3704. https://doi.org/10.3390/s22103704

Article Metrics

Article access statistics, further information, mdpi initiatives, follow mdpi.

Subscribe to receive issue release notifications and newsletters from MDPI journals

A Review of Data Centers Energy Consumption And Reliability Modeling

• November 2021
• IEEE Access PP(99):1-1

• Luleå University of Technology

Abstract and Figures

Discover the world's research

• 25+ million members
• 160+ million publication pages
• 2.3+ billion citations
• Qiankun Chang
• Yuanfeng Huang

• Angelo Algieri
• Petronilla Fragiacomo
• Ahsan Javed

• Zhonglai Wang
• Carlos Torres
• Giacomo Barbieri
• Mariela Muñoz
• Sagi Brudni
• Sapir Anidgar

• Olajide Soji Osundare
• Adebimpe Bolatito Ige

• Ming-Yen Wu
• Chung-Jhih Tsai
• Wan-Ting Su
• Shun-Hsing Hsu

• Kim Kien Nghiep Huynh

• Thai Anh Au Tran
• Hemasri Sai Lella

• Manasa Kurra

• CLUSTER COMPUT

• MULTIMED TOOLS APPL

• Hongwei Ding

• ELECTR POW SYST RES
• RELIAB ENG SYST SAFE

• Recruit researchers
• Join for free
• Login Email Tip: Most researchers use their institutional email address as their ResearchGate login Password Forgot password? Keep me logged in Log in or Continue with Google Welcome back! Please log in. Email · Hint Tip: Most researchers use their institutional email address as their ResearchGate login Password Forgot password? Keep me logged in Log in or Continue with Google No account? Sign up