Multi-track Transfer Reinforcement Learning for Power Consumption Management of Building Multi-type Air-Conditioners

Abstract

In this paper, we apply reinforcement learning to the power management control of building multi-type air-conditioners. In general, reinforcement learning requires several tens of thousands of training episodes before the control performance reaches a practical level. Therefore, applying it directly to air-conditioning control in 10-min intervals would require unrealistic training days as several years. We attempted to shorten the learning period by learning in advance on a virtual building that emulates the dynamic characteristics of an actual building. Since it is difficult to create exactly the same air-conditioning environment of the actual building, we propose a method to select the closest one from several virtual buildings based on the differences of immediate reward.