Efficient Reservoir Management through Deep Reinforcement Learning
Authors:
Xinrun Wang,
Tarun Nair,
Haoyang Li,
Yuh Sheng Reuben Wong,
Nachiket Kelkar,
Srinivas Vaidyanathan,
Rajat Nayak,
Bo An,
Jagdish Krishnaswamy,
Milind Tambe
Abstract:
Dams impact downstream river dynamics through flow regulation and disruption of upstream-downstream linkages. However, current dam operation is far from satisfactory due to the inability to respond the complicated and uncertain dynamics of the upstream-downstream system and various usages of the reservoir. Even further, the unsatisfactory dam operation can cause floods in downstream areas. Therefo…
▽ More
Dams impact downstream river dynamics through flow regulation and disruption of upstream-downstream linkages. However, current dam operation is far from satisfactory due to the inability to respond the complicated and uncertain dynamics of the upstream-downstream system and various usages of the reservoir. Even further, the unsatisfactory dam operation can cause floods in downstream areas. Therefore, we leverage reinforcement learning (RL) methods to compute efficient dam operation guidelines in this work. Specifically, we build offline simulators with real data and different mathematical models for the upstream inflow, i.e., generalized least square (GLS) and dynamic linear model (DLM), then use the simulator to train the state-of-the-art RL algorithms, including DDPG, TD3 and SAC. Experiments show that the simulator with DLM can efficiently model the inflow dynamics in the upstream and the dam operation policies trained by RL algorithms significantly outperform the human-generated policy.
△ Less
Submitted 7 December, 2020;
originally announced December 2020.
