Fusing Decision Tree and Deep Reinforcement Learning for Demand Response Optimization of Variable Volume Water Heaters

Abstract

The variable volume water heaters have greater potential to save energy and offer demand response flexibility compared to traditional fixed volume water heaters. However, optimized scheduling of variable volume water heaters must account for system uncertainties and meet the need for quick decision making. To tackle this problem, we propose a demand response optimization approach for variable volume water heaters that combines decision trees with deep reinforcement learning. Firstly, we establish the optimized scheduling framework based on fusing decision tree and reinforcement learning. Secondly, we construct a decision tree offline with training data generated by mathematical optimization methods. Then, we design a fusion model focusing on the dynamic regulation of action probability, and design a fusion algorithm and network. Simulation results show that the proposed algorithm can automatically adapt to uncertain environments and reduce energy costs of variable volume water heater by 23.2% compared to fixed volume water heaters, while satisfying user comfort requirements.