Reinforcement Learning Control for HVAC Energy Management System with Instant Operation by Selecting Virtual Building with Similar Environment

Abstract

This paper proposes a novel method for building multi-type air-conditioners, wherein the power management control system can automatically adapt to various air-conditioning environments using reinforcement learning. Our previous study reduced the learning period by pre-training on a virtual building and simulated the dynamic power characteristics of air-conditioners and room temperature. However, advantages of decreasing the learning period diminish when the difference between the virtual and actual buildings is significant. Therefore, our proposed method first performs pre-training on multiple virtual buildings with different environments. Subsequently, it selects the one whose environment is closest to that of the actual building based on the difference in average rewards.