A Graph Reinforcement Learning Algorithm for Unit Commitment

Unit commitment (UC) is the key problem in power system, and efficiently getting a solution with high quality can bring great economic and social benefits. Nowadays, as the scale of the power system has been increasing, the traditional optimization algorithms are deficient in solving accuracy and computational complexity. Artificial intelligence algorithms with powerful data process capabilities are developing rapidly and have been applied in the power system and achieved good results. Therefore, this paper proposes an algorithm to solve the UC problem based on graph reinforcement learning (GRL). Firstly, the markov decision process (MDP) of the UC problem is described based on its characteristics associated with the objective function and constraints, and the reward function is designed to update the strategy of the agent. Secondly, graph convolutional network (GCN) is chosen to better capture the topological feature of power system. Combined with deep deterministic policy gradient (DDPG) algorithm, a GRL algorithm is proposed to solve the UC problem. Thirdly, case studies are carried out to demonstrate the feasibility and performance of the proposed algorithm.