An Artificial Neural Network Approach to Affine Policy Definition in Active Distribution Networks

Legacy distribution systems need to adjust to emerging operational contexts. The rise in distributed generation will introduce efficiency challenges, and the limited communication capabilities of these systems hinder centralized control methods. Academic research has thus focused on distributed control strategies for photovoltaic generators to enhance operations and handle uncertainties. Traditional literature approaches utilize distribution system models for optimizing control parameters, which presents a computational challenge. This paper proposes a Machine Learning based agent, informed by system physics, to develop a Physics-informed neural networks environment to determine control parameters within a control structure novel from existing literature. The results suggest that the proposed methodology can define control parameters with a higher degree of flexibility, outperforming all competitive frameworks by reducing operating costs, minimizing voltage violations at 6.3%; without resorting to power spillage or increasing substation supply.