A topology-guided high-quality solution learning framework for security-constraint unit commitment based on graph convolutional network
Liqian Gao, Lishen Wei, Shichang Cui, Jiakun Fang, Xiaomeng Ai, Wei Yao, Jinyu Wen
Abstract
Security-constrained unit commitment (SCUC) is of great importance for the economic and reliable operation of the power system. The computational hardness of SCUC remains a significant issue in the power system and electricity market operations, especially with the rapid expansion of the power system, leading to increased challenges of obtaining a high-quality solution in a fast way. In this sense, this paper proposes a topology-guided high-quality solution learning framework based on graph convolutional network (GCN) and neighborhood search (NS). Firstly, a GCN-based method is presented to learn the potential patterns between commitments and graph data associated with bus feature and power grid topology. Secondly, an adaptive threshold-based method is designed to fix binary variables to achieve model reduction. Thirdly, a customized prediction-based NS is developed to restore the feasibility of the predicted commitment. Case studies with different scales verify the effectiveness and efficiency of the proposed framework for SCUC. Compared with other methods, it demonstrates the superiority of learning based on power grid graph data. In the end, it can be concluded that the feasibility and high-quality of the solution can be guaranteed while reducing most of the computation time. • A topology-guided high-quality solution learning framework is proposed. • A graph convolutional network based method for the commitment prediction is presented. • The designed adaptive threshold-based method achieves the model reduction. • The customized prediction-based neighborhood search guarantee feasibility and high quality of the solution.