Evaluation of a Machine Learning-based Algorithm for AC Optimal Power Flow

Abstract

Numerous efforts have been made to find efficient optimization methods that reduce resolution times to obtain solutions to the optimal power flow problem in alternating current (ACOPF). ACOPF is a non-convex and highly nonlinear problem. Power flow optimization problems (OPF) are usually solved using interior point methods, also known as barrier methods. One of the most commonly used approaches is the dual interior point method with filter line search. These methods are robust but expensive, as they require the calculation of the second derivative of the Lagrangian at each iteration. A promising research direction is utilizing machine learning (ML) techniques to solve operation and control problems in electrical networks. ML has been shown to significantly reduce the computational resources required in many real-world problems. Various solution methods have been employed, such as random forest, multi-objective decision tree, and extreme learning machine. In this case, ML is applied as a method that predicts voltage magnitudes and angles at each node, using physics-based network equations to calculate power injection at different nodes. For ML training, the data is divided into three sets: training, validation, and testing. These algorithms focus on minimizing their objective function and the operational cost of an AC transmission network.