Harvesting Optimal Operation Strategies from Historical Data for Solar Thermal Power Plants Using Reinforcement Learning

Operation strategy optimization for concentrating solar power (CSP) plants has been a long-studied topic in solar energy. In our work, an effective and systematic approach has been developed to harvest optimal operation strategies for CSP plants via reinforcement learning (RL). Our goal is to find the optimal strategy, which instructs the operation of different key operating variables of CSP plants in order to maximize daily power generation under a given solar irradiation series. Key variables are extracted through data mining techniques, and Deep Q-Network (DQN) is applied to find the optimal solution to the problem. A case study based on a 10MW central tower receiver solar thermal plant's operating data is carried out, which shows great improvement in power generation through our proposed method.