APPLICATION OF DEEP REINFORCEMENT LEARNING IN LOW WATER MANAGEMENT OF MULTIPLE DAMS

Abstract

 Due to climate change and severe fiscal conditions, there is a demand for the effective utilization and efficiency improvement of existing dams in dam management practices. In response, this study constructed an AI model capable of predicting appropriate discharge amounts during low water management by applying reinforcement learning techniques to three dams in the Dozan River of the Yoshino River Basin. The reinforcement learning method utilized was Twin Delayed Deep Deterministic Policy Gradient(TD3), which we believe to be more suitable for predicting dam operations compared to Deep Q-Network(DQN) method frequently used in previous research. The results showed that, compared to operations conducted in accordance with operational rules, it was possible to predict operations that could increase power generation without significantly increasing the number of drought adjustment days. This confirms the effectiveness of this method as a support for future river flow management of dams.