[3Xin4] Poster session 1

[3Xin4-78] Optimal Planning Method of Energy Plant Operation by Deep Reinforcement Learning

〇Atsuto Okada¹, Hideaki Hirano², Shoichi Kitamura², Kazuyuki Mori², Naoki Mori¹ (1.Osaka Metropolitan University, 2.Mitsubishi Electric Corporation)

Keywords:Deep Reinforcement Learning, Optimization Problem, Deep Q-Network

In recent years, energy problems have become more and more serious, and energy saving of energy plants has become an issue. Against this background, benchmarking problems for energy plant operation planning have been proposed and reported methods by mathematical optimization. In addition, deep reinforcement learning has made remarkable progress, especially in game-playing tasks. In this study, we apply Rainbow, an algorithm of deep reinforcement learning, to the benchmark problem and aim to find a solution.
Since the benchmark problem is a mixed integer nonlinear programming problem, Rainbow cannot handle continuous values of the agent's actions. In this study, we train the agents by discretizing their actions, referring to the reported solutions. As a result, some of the solutions to the benchmarking problem are as good as or better than the reference solutions. The effectiveness of the proposed method is confirmed by a
computer simulations by using the benchmark problem.