A GNN-Guided Variable Selection Approach for Efficient Derivation of the Optimal Solution in Unit Commitment

The growing scale of the power system has intensified the demand for computational efficiency in unit commitment (UC). This paper introduces a novel approach that addresses this issue by adopting graph neural networks (GNN) into the Branch and Bound (B&B), integrating the characteristics of variable selection (VS) within UC problems. Our method focuses on efficiently obtaining optimal solutions using the GNN-guided VS approach. Specifically, we investigate the characteristics of the high-quality but time-consuming VS heuristic when solving UC problems. Then, the GNN incorporates these characteristics to make high-quality VS decisions accurately and quickly. In the implementation, the UC model is solved using a high-quality but time-consuming VS heuristic to collect training data. These data are employed to train the GNN model with designed label-smooth and depth-adaptive mechanisms, which exploit these characteristics in UC problems for acceleration. To evaluate the effectiveness, we conduct various experiments on the IEEE118, RTS-GMLC, and a real-world provincial system in China. The overall performance demonstrates significant improvements (4.6x, 1.5x, and 1.8x, respectively) in computational efficiency. The comparison experiments show that the proposed mechanisms contribute to enhanced learning performance and, thus, further UC acceleration. Moreover, our approach exhibits promising generalization capabilities.