Optimal Power Flow With Physics-Informed Typed Graph Neural Networks

This work describes a new way to solve the optimal power flow problem applying typed graph neural networks. Typed graph neural networks allow the representation of different elements of transmission systems with different types of nodes, thus adding accuracy and interpretability to the solutions obtained, in comparison to results obtained with conventional feed-forward neural network models. The proposed graph neural network architecture is trained without the need of training data, through a physics informed loss function which incorporates not only the optimization objective, but also operational constraints of the physical system. Results are comparable with those obtained with the interior point method, and it is shown that the calculation time is greatly reduced.