EVALUATION OF DEEP REINFORCEMENT LEARNING-BASED DAM OPERATION MODEL WITH UNCERTAIN INFLOW PREDICTION

Abstract

The effect of the inflow prediction error was investigated in a dam operation model, which is based on the neural network with reinforcement learning. The dam operation model used in the study determines the appropriate dam discharge according to the situation that changes from moment to moment (reservoir water level up to the present time, observed inflow and timeseries of predicted inflow up to 6 hours later). The examination object was Matsubara Dam of the Chikugo River system. For model training, virtual flood data created by extending historical rainfall was used, and reinforcement learning (Deep Q Learning) was applied. In order to verify the model, artificial error was added to the inflow of virtual dam inflow data, and used as the virtual inflow prediction data during actual operation. In the verification with three virtual floods exceeding the design flood scale, the model showed a reasonable operational judgment and hardly affected by the error of the predicted inflow.