Abstract: This paper focuses on the key issue of the Three Gorges hydropower station’s in-plant economic operation, which is aimed at achieving a real-time load allocation of large-scale units for minimizing water consumption. Dynamic programming usually encounters the curse of dimensionality when dealing with a large-scale hydropower unit cluster, and therefore, it cannot meet the requirement of real-time dispatching decision for the station. For training a multi-period unit load distribution model and its decision-making, we develop a deep reinforcement learning-based framework to train the deep neural network and generates unit load distribution plans through a pre-trained network model. We apply a group theory idea to processing the state and action features of the learning, so as to compress the state and action space significantly and improve model training efficiency. The results indicate that compared to dynamic programming, our new method shortens the decision-making time by two orders of magnitude at a cost of less than 1% benefit loss. Thus, it offers a rapid and efficient solution for the unit load allocations in large-scale hydropower stations.

Key words: economic operation, unit commitment, deep reinforcement learning, real-time decision