Learning to Predict Security Constraints for Large-Scale Unit Commitment Problems

Ahead-of-time electricity generation and consumption scheduling, also known as unit commitment, is essential to operate power grids. Today, it is usually formulated and solved as a mixed-integer program. To ensure robustness against operational contingencies, a large number of security constraints must be considered, which significantly increases the problem’s complexity. However, only a small subset of these constraints is typically active in the solution. Conventional solving approaches attempt to eliminate redundant security constraints through a computationally expensive iterative search. In this paper, we study machine learning approaches to alleviate this search by predicting a set of relevant security constraints to retain. We propose a new neural network architecture based on graph convolutions and the attention mechanism, which we evaluate on synthetic unit commitment problems for nine power grids. Compared to the conventional iterative approach and previous machine learning-based methods, our architecture improves average and (more importantly) worst-case solving times by a factor of 2 to 3.