Abstract Body

Data centres (DCs) energy consumption issues have largely aroused worldwide concerns. There are considerable studies that investigate the areas on a configuration and building designs, or location-based level to save the energy consumption of the DC. However, changing the configurations or moving DC is relatively problematic. It has been recognised that there exists a trade-off relationship between the IT and the cooling energy consumption in the DC as keeping the lower temperature in the IT room would increase the energy consumed by the cooling devices but at the same time would ensure the IT computational capability, therefore, reduced the IT energy consumption, and vice versa. However, The literature lacks a sophisticated analysis of how to utilise the existing facilities in DC to minimise energy consumption by active intervention. To verify this trade-off relationship, we propose a managerial strategy that aims to optimise the DC energy consumption by controlling the combinations of Air-Conditioners (ACs) target temperatures. The enumerated existing solutions to similar questions have dropped in the fields of pure mathematical modelling or non-model-based machine learning, which have the limitations of consuming overmuch human and computational effort. Differ from the optimisation solutions to similar questions such as pure mathematical modelling, dynamic programming (DL), or Artificial Neural Network (ANN) and Reinforcement Learning (RL), we adopted the joint approach Statistical Formulated Reinforcement Learning (SFRL) based on a multivariate Vector Auto-regressive Model (VAR). VAR as a multivariate time series model formulated by analysing the real-time historical data collected from a DC locates in Turkey, and act as an approximation function to the optimisation solver set up by RL approach. This largely solved the difficulties both in the analytical modelling and machine learning fields as it reduced the manual effort of modelling the comprehensive environment by mathematical approaches as well as the computational effort of storing the infinite optimal strategies when RL makes the optimal step-solution. Our contribution to the literature mainly about the application of functional approximation to the RL decision process and its practical optimisation usage in the industry.