Learning to Configure Hyperparameters in Solving Unit Commitment Problems with Mixed Integer Linear Programming Solvers

Unit commitment problems can be solved more efficiently with mixed integer linear programming solvers when more preferred hyperparameters are configured. We propose a learning approach to configure hyperparameters automatically for unit commitment problems. Our method employs a bipartite graph to represent the underlying structure of MILP and leverages a Graph Convolutional Network to adjust solver parameters on an instance-by-instance basis. The effectiveness of this approach is empirically validated through a series of numerical experiments, which demonstrate significant improvements in solver performance, resulting in noticeable computational time savings. These findings highlight the practicality and potential of our strategy in real-world applications of unit commitment, suggesting a promising new direction for optimization solver configuration.