Task Scheduling for Energy Efficiency in Data Center: Machine Learning Approach

Energy efficiency in Data Center has becoming crucial issue in recent decades, considering the high energy consumption in Data Centers along with the development of digital era. Cooling system is a part of Data Center with highest energy consumption. This fact explains the need for long-lasting systems that manage energy consumption while ensuring optimal system performance. The aim of this paper is to develop and design a Task Scheduling to improve energy efficiency using machine learning approach. The goal is to maintain temperature in the range of 22°C - 26°C and reduce Power Usage Effectiveness (PUE) below 1.8. To support this goal, machine learning is used to obtain temperature prediction values. This research compares three types of algorithms: Long Short-Term Memory (LSTM), Autoregressive Integrated Moving Average (ARIMA), and Prophet to get the best prediction result. LSTM shows the lowest Root Mean Squared Error (RMSE) value (0.1617720) compared to ARIMA (0.111928) and Prophet (0.087384). It can be concluded that LSTM algorithm give more accurate prediction than the others. The task scheduling results then creates 24 best actions (schedules) that used as a reference for agents to carry out simulations in December to maximize the rewards that will be obtained for 30 days or 30 episodes per 24 actions/episodes. By using those methods, it can reduce the PUE become 1.76 from 1.9 and the temperature below 26°C. It means that the using of machine learning to reduce PUE by creating best schedule action space can increase the energy efficiency (DCiE) by 4%.