OPF-HGNN: Generalizable Heterogeneous Graph Neural Networks for AC Optimal Power Flow

The precise solution of the Alternating Current Optimal Power Flow (AC-OPF) problem is a pivotal challenge in the domain of real-time electricity grid operations. This problem is notorious for its significant computational complexity, primarily attributable to its inherently nonlinear and nonconvex nature. Recently, there has been a growing interest in harnessing Graph Neural Networks (GNN) as a means to tackle this optimization task, leveraging the incorporation of grid topology within neural network models. Nonetheless, existing techniques fall short in accommodating the diverse array of components found in contemporary grid networks and restrict their scope to homogeneous graphs. Furthermore, the constraints imposed by the grid networks are often overlooked, resulting in suboptimal or even infeasible solutions. To address the generalization and effectiveness of existing end-to-end OPF learning solutions, we propose OPF-HGNN, a new graph neural network (GNN) architecture and training framework that leverages heterogeneous graph neural networks and incorporates the grid constraints in the node loss function using differentiable penalty regularization. We demonstrate that OPF-HGNN is robust and outperforms traditional GNN learning by two orders of magnitude traditional GNN learning across a large variety of real-world grid topologies and generalization settings.