Machine learning applications in data centre cooling performance optimisation and forecasting: A review

The demand for data centres (DCs) is surging due to the explosive growth of cloud services and social media. In 2022, DCs accounted for 1% of global electricity consumption, while in the United Kingdom, this figure further increases to 2.5%. Even though the IT system attributes primarily (around 56%) to a DC’s total power consumption, depending on different cooling techniques (air, liquid, and free cooling), the cooling system alone can make up to 30% to 40% of the energy demand. As a subset of free cooling, dew-point evaporative cooling (DPC) can achieve a Coefficient of Performance (COP) of up to 52.5, compared to the COP of 2 to 4 of most wildly applied traditional mechanical vapour compression cooling. Therefore, applying DPC to DC can achieve significate energy saving and carbon reduction. To meet the cooling demand of DCs without sacrificing thermal safety, preliminary modelling is required during the designing stage. While due to the dynamics natural of DCs cooling system, it often results in inadequate optimisation and inaccurate forecasting of cooling performance which require hand-tuning for specific devices. In this case, machine learning (ML) technique fundamentally outperform conventional modelling as it only learns and derives from the massive data collected from the desired system. Combining ML technique with DCs cooling can achieve energy saving for up to 26.6%, 15%, and 72% of air, liquid and dew-point cooling respectively under diverse climate conditions, which makes integrating ML into DC cooling systems, especially DPC, a promising research area. This review paper focuses on the development of ML techniques in DCs’ control strategy optimisation and long-term performance forecasting from 2000 to 2022. Existing literatures are categorized based on seven elements, while five ML based optimisation strategies and six forecasting methods are firstly introduced, followed by jointly optimizing approaches with IT system. Based on above analysis, existing research gaps in DC cooling optimisation and forecasting are identified, and associated recommendations are offered from the standpoint of the authors. The findings of this review can act as a steppingstone for scholars, researchers and policy makers in related fields, therefore contribute to the advancement of the SET agenda.