Primal heuristics for Dantzig-Wolfe decomposition for unit commitment

TL;DR

This paper introduces two primal heuristics for Dantzig-Wolfe decomposition in unit commitment problems, one based on decomposition and the other on machine learning, to efficiently find near-optimal solutions.

Contribution

It proposes novel primal heuristics, including a machine learning approach, for improving solution efficiency in Dantzig-Wolfe decomposition for energy unit commitment.

Findings

ML-based heuristic performs better at higher suboptimality tolerances.

Decomposition-based heuristic is more accurate at lower tolerances.

Both heuristics outperform solving the original problem directly.

Abstract

The unit commitment problem is a short-term planning problem in the energy industry. Dantzig-Wolfe decomposition is a popular approach to solve the problem. This paper focuses on primal heuristics used with Dantzig-Wolfe decomposition. We propose two primal heuristics: one based on decomposition and one based on machine learning. The first one uses the fractional solution to the restricted master problem to fix a subset of the integer variables. In each iteration of the column generation procedure, the primal heuristic obtains the fractional solution, checks whether each binary variable satisfies the integrality constraint and fix those which do. The remaining variables are then optimised quickly by a solver to find a feasible, near-optimal solution to the original instance. The other primal heuristic based on machine learning is of interest when the problems are to be solved repeatedly with different demand data but with the same problem structure. The primal heuristic uses a pre-trained neural network to fix a subset of the integer variables. In the training phase, a neural network is trained to predict for any demand data and for each binary variable how likely it is that the variable takes each of two possible values. After the training, given an instance to be solved, the prediction of the model is used with a rounding threshold to fix some binary variables. Our numerical experiments compare our methods with solving the undecomposed problem and also with other primal heuristics from the literature. The experiments reveal that the primal heuristic based on machine learning is superior when the suboptimality tolerance is relatively large, such as 0.5% or 0.25%, while the decomposition is the best when the tolerance is small, for example 0.1%.