Accelerating L-Shaped Two-Stage Stochastic SCUC With Learning Integrated Benders Decomposition

Benders decomposition is widely used to solve large mixed-integer problems. This article takes advantage of machine learning and proposes a variant of Benders decomposition to tackle two-stage stochastic security-constrained unit commitment (SCUC). The problem is decomposed into a master problem (MP) and subproblems (SPs) corresponding to individual load scenarios. The primary objective is to mitigate computational expenses and memory consumption associated with Benders decomposition by generating tighter cuts and reducing the MP's dimensions. A regressor reads load profile scenarios and predicts objective function proxy values for the SPs, enabling the creation of tighter cuts for the MP. The numerical difference between cut values and proxy variable values serves as the basis for identifying useful cuts. Analytical cut-filtering and classification-assisted cut-filtering approaches are discussed and compared. Useful cuts contain the necessary information to form the feasible region and are iteratively added to the MP, whereas nonuseful cuts are discarded, thus reducing the computational burden at each Benders iteration. Simulation studies conducted across various test systems demonstrate the efficacy of the proposed learning-enhanced Benders decomposition in solving two-stage SCUC problems, showcasing superior performance compared to conventional multicut Benders decomposition and offering numerical advantages over cut classifier-based Benders approaches.