Fast Unit Commitment Constraint Screening with Learning-Based Cost Model

Xuan He, Honglin Wen, Yufan Zhang, Yize Chen, Danny H. K. Tsang

Abstract: Unit commitment (UC) is one of the cornerstone problems for power system operators in finding the most economical and feasible generation schedules and power dispatch. Yet with ever-increasing complexities coming from both generation and demand sides, it becomes challenging to find UC solutions in a timely manner with huge combinatorial problems involved. Constraint screening holds the promise of reducing the number of redundant or inactive constraints in the original UC problem so that the solution process can be accelerated by solving a smaller problem. In this paper, we bridge the strong representation capability of machine learning (ML) models to the usually cumbersome constraint screening procedure. By a novel learning paradigm to estimate the most economical costs given load profiles and guiding the screening procedure, our approach can screen out a higher proportion of operational constraints. We verify the proposed method’s performance for both sample-aware setting (a specific load vector) and sample-agnostic setting (a given region of load) on a variety of UC setups. In the former 93.4% of redundant constraints are screened out, while we reduce over 64% of solution time in the latter case.