Optimal Active and Reactive Power Dispatch of Distributed Generators Using an Input Convex Neural Network Surrogate Model

With the increasing demand for clean and sustainable energy, there is a growing push to integrate Renewable-based Distributed Generators (RDGs) into distribution networks (DNs). While RDGs offer significant environmental benefits, improper control and coordination can result in poor voltage regulation, increased losses, and system instability due to their non-dispatchable and intermittent nature. To ensure reliable and efficient integration, this work presents an optimization strategy to optimally coordinate two or more RDGs to dispatch both active and reactive power. Due to the nonconvex and nonlinear nature of AC power flow equations, they pose a computational challenge, with no guarantee of obtaining optimal solutions using gradient-based optimization techniques. This work proposes the use of data-driven surrogate models for optimization. A surrogate model is developed by training a neural network to estimate system outputs from given inputs. The primary objective is to obtain optimal dispatch signals for the RDGs to minimize average voltage deviation and %losses across the network while satisfying operational constraints, including voltage limits, current ratings, and voltage unbalance factor. The proposed approach was bench-marked against models trained with other activation functions, as well as a metaheuristic method that uses the actual system model for optimization. Results demonstrate that the Input Convex Neural Network (ICNN) surrogate, combined with a nonlinear solver, efficiently evaluates the optimal values, substantially reducing computation time, while maintaining constraint compliance and delivering reliable dispatch signals.