Fast Solution of Unit Commitment Using Machine Learning Approaches

The complexity of energy scheduling problems is increasing due to the energy transition. In recent research, Machine Learning (ML) has shown potential to contribute to the methodology for executing these tasks efficiently and reliably in future. This paper develops and compares three approaches for predicting binary decisions in Unit Commitment problems with network constraints: Two ML predictors using Random Forests and Graph Neural Networks are contrasted with a rule-based approach. On large datasets of realistic synthetic Unit Commitment problems, the performance criteria that need to be met for successful real-word application are evaluated: What is the speedup potential of using the predictions in the process? What is the risk of losing optimality or even feasibility? And what are the generalization capabilities of the predictors? We find that all three approaches have promising potential, each approach having its own pros and cons.