A Deep Reinforcement Learning Algorithm for Accelerating Unit Commitment Optimization

Unit commitment (UC) is a critical issue in the optimal scheduling of power systems. As the number of units increases, the scale of the problem expands rapidly, making it a significant challenge to solve the UC problem quickly. This paper addresses the inefficiency of UC model solving by innovatively combining power systems with Graph Neural Networks and Reinforcement Learning. First, a Markov decision chain model is established for the unit commitment problem. Then, an Edge-Conditioned Convolution Network - Long Short-Term Memory - Twin Delayed Deep Deterministic Policy Gradient (ECT-TD3) model is constructed to quickly obtain an initial solution for the unit commitment problem. Finally, a feasibility recovery verification is performed, followed by the use of a solver to obtain the final solution rapidly. System case studies conducted on data generated from real-world load scenario characteristics demonstrate that the proposed framework effectively accelerates the solution of the UC problem, increasing the solving speed by about three times.