Machine learning approaches to the unit commitment problem: Current trends, emerging challenges, and new strategies

Abstract

Traditional power system operation and control decision-making processes, such as the unit commitment (UC) problem, primarily rely on the physical models and numerical calculations. With the growing scale and complexity of modern power grids, it becomes more complicated to accurately formulate the physical power system and more difficult to efficiently solve the corresponding UC problems. As a matter of fact, plenty of historical power system operation records as well as real-time data could provide useful information and insights of the underlying power grid. To this end, machine learning methods could be valuable to help understand the relationship of UC performance to power system parameters, reveal the rationality behind such relationship, and finally address UC problems in a more efficient and accurate way. This article discusses the current practices of using machine learning approaches to solve the mixed-integer linear programming based UC problems. The associated challenges are analyzed, and several promising strategies for adopting machine learning approaches to effectively solve UC problems are discussed in this article. In addition, we will also explore machine learning approaches to promptly solve steady-state nonlinear AC power flow and dynamics differential equations, so that they can be integrated into the UC problems to guarantee AC power flow security and dynamic stability of system operations, as compared to the current DC power flow constrained UC practice. Our studies show that machine learning, as model-free methods, is a valuable alternative or addition to the existing model-based methods. As a result, the effective combination of machine learning based approaches and physical model based methods are expected to derive more efficient UC solutions that can improve the secure and economic operation of power systems.