Data Centres (DCs), the core of the ever-increasing economic and societal activities, are experiencing high energy consumption due to the rapidly growing demand for digital services, which leads to the largest DCs' operational costs and significant environmental and power security impacts. Hence, optimising their energy efficiency is a critical and top priority to ensure economic and environmentally sustainable DC management. However, prior heuristics and engineering-based solutions are inadequate due to the increasing physical complexity and sheer number of possible configurations, non-linear system interactions, and the growing monitoring of operational management data. Therefore, this paper develops a six-layered hybrid Convolutional Neural Network (CNN) and Long-Short-Term Memory (LSTM), called CNN-LSTM, a suitable data-driven Deep Learning (DL) model inspired by the human brain for effectively modelling the complex and plant performance and optimising DC efficiency. The proposed model uses CNN to capture complex parameter interactions and spatial pattern recognition and LSTM to capture temporal dependencies of the time series operational data to predict the next energy consumption value and optimise it using a cooling system fan speed controllable variable. The model was extensively trained and tested using actual operational management data obtained from the Enea High-Performance Computing (HPC) CRESCO6 cluster, Italy. 80% of the data was used for training, and the rest 20% for testing. According to the experimental results, it accurately predicts energy consumption every 15 minutes with an average Mean Absolute Error (MAE) of 0.0043. A sensitivity analysis optimisation strategy is also implemented using various cooling system fan speed set points. When the fan speed optimal set-point is automatically reduced by 50%, the energy consumption is optimised by an average of 0.0029 kWh with a 0.0039 MAE over the last 15 minutes of the testing set. Hence, this paper found that the proposed CNN-LSTM deep learning predictive model can effectively mimic actual DC operations and optimise efficiency. By simulating this model, DC operators can effectively manage and optimise their DC energy efficiency while reducing energy and environmental costs.
A Hybrid Deep Learning Approach for Modelling and Optimising Data Centre Energy Efficiency
De Chiara D.;Chinnici M.
2024-01-01
Abstract
Data Centres (DCs), the core of the ever-increasing economic and societal activities, are experiencing high energy consumption due to the rapidly growing demand for digital services, which leads to the largest DCs' operational costs and significant environmental and power security impacts. Hence, optimising their energy efficiency is a critical and top priority to ensure economic and environmentally sustainable DC management. However, prior heuristics and engineering-based solutions are inadequate due to the increasing physical complexity and sheer number of possible configurations, non-linear system interactions, and the growing monitoring of operational management data. Therefore, this paper develops a six-layered hybrid Convolutional Neural Network (CNN) and Long-Short-Term Memory (LSTM), called CNN-LSTM, a suitable data-driven Deep Learning (DL) model inspired by the human brain for effectively modelling the complex and plant performance and optimising DC efficiency. The proposed model uses CNN to capture complex parameter interactions and spatial pattern recognition and LSTM to capture temporal dependencies of the time series operational data to predict the next energy consumption value and optimise it using a cooling system fan speed controllable variable. The model was extensively trained and tested using actual operational management data obtained from the Enea High-Performance Computing (HPC) CRESCO6 cluster, Italy. 80% of the data was used for training, and the rest 20% for testing. According to the experimental results, it accurately predicts energy consumption every 15 minutes with an average Mean Absolute Error (MAE) of 0.0043. A sensitivity analysis optimisation strategy is also implemented using various cooling system fan speed set points. When the fan speed optimal set-point is automatically reduced by 50%, the energy consumption is optimised by an average of 0.0029 kWh with a 0.0039 MAE over the last 15 minutes of the testing set. Hence, this paper found that the proposed CNN-LSTM deep learning predictive model can effectively mimic actual DC operations and optimise efficiency. By simulating this model, DC operators can effectively manage and optimise their DC energy efficiency while reducing energy and environmental costs.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
