A Fast-Fixing Method for Stochastic Unit Commitment: A Trade-Off Analysis Between Accuracy and Efficiency

Abstract

Stochastic unit commitment (SUC) addresses the uncertainties associated with renewable energy resources by generating various scenarios, which challenges the solution efficiency due to the vast number of variables and constraints involved. Existing methods apply machine learning techniques to accelerate the process by directly predicting generator outputs. However, machine learning may not deliver feasible solutions that meet all the constraints. This paper proposes a fast-fixing method (FFM) that predicts the probability distribution of unit statuses instead of directly predicting the dispatch decisions. With units on/off statuses partly prefixed, the original large-scale mixed-integer linear programming (MILP) problem will be simplified. The number of prefixed units can be adjusted by a threshold predicting acceptance (PA), providing a flexible choice for users to balance the trade-off between efficiency and accuracy. According to the case study on RTS-96 involving 15 to 30 generator units and 20 scenarios, when setting PA to 0.1, the proposed FFM significantly decreases the average solving time by 45.28% compared to CPLEX, 28.6% compared to Gurobi, and 21.43% compared to Benders decomposition, while only sacrificing a uniform 0.34% in average accuracy across all comparisons.