Enhanced Optimal Power Flow Using a Trained Neural Network Surrogate for Distribution Grid Constraints

A team of researchers from the University of Cyprus and Imperial College London has published a novel framework that uses a trained neural network as a surrogate to dramatically speed up critical power grid calculations. The paper, 'Enhanced Optimal Power Flow Using a Trained Neural Network Surrogate for Distribution Grid Constraints,' addresses a major bottleneck in managing modern electricity grids. As distributed energy resources like rooftop solar, electric vehicles, and heat pumps proliferate, grid operators need fast, accurate tools to ensure stability—a problem known as Optimal Power Flow (OPF). Traditional OPF models are computationally heavy and often rely on approximations that can fail. This new method trains a neural network to learn the complex relationship between power flow and voltage, then encodes that network's exact input-output map into the optimization problem using mixed-integer linear programming (MILP).

This approach guarantees a globally optimal solution within the solver's tolerance, a significant advantage over approximate methods. The researchers validated their NN-OPF on a realistic low-voltage network model integrating photovoltaic generation, EV charging, and heat pumps. The results showed exceptional accuracy, with maximum voltage deviations of less than 1.0 Volt during validation. Crucially, the computation time was 'substantially reduced' compared to full nonlinear OPF models, making it competitive with other fast approximations like Second-Order Cone Programming (SOCP) DistFlow, but with higher fidelity. This breakthrough paves the way for near-real-time grid management and more reliable integration of renewable energy, moving AI from a predictive tool to an embedded component of core grid optimization.