Unit Commitment Warm-Start Decision Method Based on Similarity Computation via Graph Convolutional Networks

With the rapid development of AI, the ability to make fast and accurate unit commitment (UC) decisions is of great significance. This paper proposes a hybrid UC decision model based on both physical and data-driven approaches, and develops a UC warm-start decision method based on Similarity Computation via Graph Convolutional Networks (SIMGCN). Firstly, we design the labels for UC scheme similarity scores considering the weighted rated capacity of units. Then, we construct a graph similarity evaluation model based on Graph Convolutional Networks (GCN). Leveraging the advantages of GCN in establishing power grid topology and utilizing bus-level information, we train a mapping model between bus load and graph similarity scores using historical data. Finally, we input the decision scheme with the highest historical similarity as the initial feasible solution into the physical model to obtain the optimal UC solution for the day to be predicted through warm-start. Simulation results based on the IEEE-118 bus system show that compared to traditional physical-driven decision methods, the proposed method has superior solving time while ensuring decision accuracy.