GNN-Based Approximation of AC-OPF for Reactive Power Management at the Transmission-Distribution Interface

Abstract

The increasing penetration of distributed energy resources (DERs), the growing usage of power electronic devices, and the extensive replacement of overhead lines by underground cables have made the reactive power management at the transmission-distribution (T-D) interface more challenging. Alternating current optimal power flow (AC-OPF) provides physically feasible reactive power dispatch decisions for controllable resources. However, long-term horizon studies require solving the AC-OPF problem repeatedly, which becomes computationally demanding. To address this challenge, this paper proposes a graph neural network (GNN)-based approximation of AC-OPF for reactive power management. AC-OPF generates reference reactive power setpoints for a set of controllable photovoltaic (PV) units, which form the supervised-learning labels. Based on the hourly operational data collected over four months, a graph attention network (GAT) is trained to predict the target vector from graph-structured operating states. Performance is assessed using setpoint-approximation error, feasibility after applying the predicted actions and resolving load flow, and computational runtime compared with repeated AC-OPF.