# Use of Machine Learning Models to Warmstart Column Generation for Unit Commitment

arXiv: https://arxiv.org/abs/2110.06872 · HTML: https://mlsanity.com/paper/2110.06872/

## Key facts

- Cited by 5 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 13 Oct 2021.

## Built on by (most cited first, 5 of 5)

- [Dual Lagrangian Learning for Conic Optimization](https://mlsanity.com/paper/2402.03086/) (2024)
- [Predicting Accurate Lagrangian Multipliers for Mixed Integer Linear Programs](https://mlsanity.com/paper/2310.14659/) (2023)
- [Towards An Unsupervised Learning Scheme for Efficiently Solving Parameterized Mixed-Integer Programs](https://mlsanity.com/paper/2412.17623/) (2024)
- [Learned Pairwise Deep Dual-Optimal Inequalities for Stabilizing Column Generation](https://mlsanity.com/paper/2607.13373/) (2026)
- [Memory-Computation Tradeoffs in Semi Amortized Parametric Optimization](https://mlsanity.com/paper/2607.20769/) (2026)

## Abstract (arXiv)

The unit commitment problem is an important optimization problem in the energy industry used to compute the most economical operating schedules of power plants. Typically, this problem has to be solved repeatedly with different data but with the same problem structure. Machine learning techniques have been applied in this context to find primal feasible solutions. On the other hand, Dantzig-Wolfe decomposition with a column generation procedure has been shown to be successful in solving the unit commitment problem to tight tolerance. We propose the use of machine learning models not to find primal feasible solutions directly but to generate initial dual values for the column generation procedure. Our numerical experiments compare machine learning based methods for warmstarting the column generation procedure with three baselines: column pre-population, the linear programming relaxation and coldstart. The experiments reveal that the machine learning approaches are able to find both tight lower bounds and accurate primal feasible solutions in a shorter time compared to the baselines. Furthermore, these approaches scale well to handle large instances.