Randomized Branching Strategy in Solving SCUC Model

Abstract

The Branch-and-Bound (BnB) method is the fundamental solution framework for solving large-scale security-constrained unit commitment (SCUC) problem. Due to the central role variable selection rules play in such a solution procedure, this paper develops some efficient methods to actively learn the variable selection rule. Instead of using a pre-fixed rule, we propose to use a randomized strategy to select the branching variables. In such a strategy, the probability associated with the variable selection is learned from the historical solutions generated by the similar problems with different parametric patterns. To accelerate the learning procedure, we further propose to use either Grid Search or Bayesian Optimization technique to learn such a probability distribution. Using the randomly generated SCUC problems, we evaluate our randomized variable selection rule which incorporates the Most Infeasible Branching rule, Least Infeasible Branching rule, Pseudocost Branching rule, and the CPLEX Adaptive Branching rule as a basis. The preliminary computational results show that our proposed method gives remarkable improvements in solving the SCUC problem.