Machine learning-accelerated distributed optimisation methods for optimal power flow: A review

Abstract

Modern power grids have become increasingly complex, with greater uncertainty due to the widespread integration of renewable energy resources potentially leading to higher operating costs. The optimal operation of these networks can be accomplished using optimal power flow (OPF), a fundamental optimisation tool for power networks with objectives including generation cost minimisation. Whilst the OPF problem itself is not new, quickly solving problems of a practical scale remains an active research area. Two approaches here are distributed optimisation and, more recently, machine learning (ML). Distributed optimisation improves scalability, avoids single points of failure, and enhances user privacy, whilst ML has the potential to provide solutions significantly faster than traditional optimisation methods. The goal of this review is to present approaches that overlap both areas, identifying complementary aspects as well as areas for further exploration. For example, one drawback of the alternating direction method of multipliers (ADMM), a distributed optimisation algorithm, is that it has slow convergence. Several reviewed papers have mitigated this by using ML to accelerate convergence through the prediction of consensus variable values, demonstrating improvements in terms of convergence time. Challenges remain, including the generalisation of results across different network topologies, something with the potential to be addressed with additional ML models such as graph neural networks (GNNs). Further areas to explore at the intersection of these two fields are identified, including augmented Lagrangian alternating direction inexact Newton (ALADIN) and overlapping Schwarz decomposition optimisation methods, as well as ML models such as GNNs and physics-informed neural networks (PINNs).