Analyzing Data Characteristics for Learning OPFs

Renewable energy sources have altered the generation and loading characteristics of modern electrical power systems, which is why finding the solution to the Power Flow Problem (OPF) has increased in complexity and the need to solve it more frequently. Despite increasing complexity, the OPF has a known structure presenting features that allow the use of Machine Learning (ML) to find solutions. This paper sets out to investigate the effects of varying specific characteristics of the input data used to train ML models for solving the OPF problem, particularly the load. We examine how the sample size, load correlation coefficient and maximum load deviation impact the performance of a neural network trained to predict OPF solutions. The results of this research demonstrate a clear correlation between these variations in the training dataset and the performance of different ML models.