The optimised operation of cooling devices in data centres: A VAR-PSO-based MPC scheme

Abstract

Data centres (DCs) are the most significant energy consumers globally, where IT and cooling devices account for approximately 45% and 55% of their total energy consumption, respectively. Despite the extensive research conducted on reducing the energy consumption of IT devices, studies focusing on the reduction of energy consumption of cooling devices in DCs are relatively limited. Furthermore, there is a lack of research on the optimal utilisation of existing cooling devices to minimise their energy consumption. In this study, a Model Predictive Control (MPC) framework, in which a Vector Autoregressive Model (VAR) and Particle Swarm Optimisation (PSO) are integrated, is proposed to optimise the energy consumption in DCs by controlling the temperature setpoints of air conditioners (ACs). The VAR model is employed to capture the causal-effect relationships among the system variables, which affect the temperature changes in DC rooms, and then used to predict future temperature parameters over time. The PSO algorithm is utilised to find the optimal temperature setpoints combinations based on the future temperature changes provided by the VAR model. The Model Predictive Control (MPC) framework controls the VAR model and the PSO optimisation, continuously evaluating the optimised operation of air conditioners (ACs) and adjusting the VAR model if any deviation is detected to ensure the energy efficiency of DCs falls within the predefined range. Through this approach, the optimisation problem can be solved dynamically, taking future performance into considerations, and proactively avoiding potential issues. The feasibility of the proposed MPC framework has been tested in a national DC room in Thailand. Moreover, the effectiveness of the framework under various operating scenarios was validated through a Computational Fluid Dynamic (CFD) simulated environment. The results of the field experiment demonstrate that the proposed MPC framework is effective in reducing energy consumption in the DC room, achieving a 32.5% reduction compared to existing cooling practices that utilise fixed AC setpoints during operation. Additionally, the simulation results illustrate a high adaptability of the proposed approach to changing conditions. This study makes significant contributions at the intersection of theoretical, methodological, and practical domains. In terms of theoretical contributions, this study challenges prevailing paradigms in DCs’ energy optimisation. By emphasizing managerial solutions over traditional hardware modifications, our approach offered a novel perspective on effective energy optimisation strategies. Methodologically, our research introduces a dynamic framework (MPC) integrating predictive modelling (VAR) and optimisation algorithms (PSO). The integration of VAR, PSO, and MPC approaches not only leverages their respective strengths but also compensates for their individual limitations, maximising the synergistic potential. Our method also provides a flexible and adaptive solution by autonomously adjusting to changing system states. This adaptive quality bridges the gap between theory and practical application, differentiating our approach significantly from conventional practices in comparable industries. Practically, the proposed approach has been proved effective in controlling the temperatures in the DC room, achieving notable energy savings. Furthermore, the experiment demonstrates that the proposed MPC framework responds to workload changes within a reasonable timeframe, indicating its’ real-time adaptability.