Is learning for the unit commitment problem a low-hanging fruit?

TL;DR

This paper demonstrates that simple, naive algorithms can efficiently solve the NP-hard unit commitment problem in power systems, challenging the necessity of complex learning-based methods.

Contribution

The study shows that straightforward algorithms can achieve near-optimal solutions quickly, questioning the added value of complex machine learning approaches for UCP.

Findings

Naive algorithms find optimal or near-optimal solutions efficiently.

Speedups are significant compared to traditional methods.

Complex learning methods must demonstrate clear improvements to justify their complexity.

Abstract

The blast wave of machine learning and artificial intelligence has also reached the power systems community, and amid the frenzy of methods and black-box tools that have been left in its wake, it is sometimes difficult to perceive a glimmer of Occam's razor principle. In this letter, we use the unit commitment problem (UCP), an NP-hard mathematical program that is fundamental to power system operations, to show that simplicity must guide any strategy to solve it, in particular those that are based on learning from past UCP instances. To this end, we apply a naive algorithm to produce candidate solutions to the UCP and show, using a variety of realistically sized power systems, that we are able to find optimal or quasi-optimal solutions with remarkable speedups. Our claim is thus that any sophistication of the learning method must be backed up with a statistically significant improvement of the results in this letter.