New Hybrid Neural Network Architecture Improves AC Optimal Power Flow Solutions

A Hybrid Heterogeneous Message Passing Neural Network (HH-MPNN) has been developed to address computational challenges in AC Optimal Power Flow (ACOPF) for large-scale electrical grids. Traditional solvers often require excessive time for these complex calculations. Machine learning approaches previously faced limitations in scalability and adaptability to different grid configurations. The HH-MPNN architecture combines a heterogeneous graph neural network with a transformer component and physics-informed positional encodings. This design allows modeling of distinct power system elements while capturing both local characteristics and long-range dependencies through global attention mechanisms. Testing on PGLearn and GridFM-DataKit benchmarks demonstrated the model's effectiveness across systems ranging from 14 to 2,000 buses. For standard grid topologies, the approach achieved optimality gaps below 1%. The architecture also showed strong performance in N-1 contingency scenarios, demonstrating zero-shot generalization capabilities with optimality gaps under 3%. This research addresses critical limitations in existing machine learning methods for power system optimization.