FLOOD CONTROL BY THE DAM OPERATION MODEL USING DEEP REINFORCEMENT LEARNING

Abstract

We developed the deep neural network model of dam operation for flood control. The input of the model was observed and predicted dam inflow, current dam discharge and dam water level. The output of the model was the dam discharge at the next time step. Deep reinforcement learning was applied to train the network. First, lots of virtual floods were prepared by flood simulation model. Then, the network was trained so as to appropriately control the floods. Developed model was applied to the Matsubara dam in Chikugo River. The model result was compared with the operation rules, and consequently validity was confirmed.