Augmented Lagrangian Guided Learning for the Optimal Power Flow Problem

This paper investigates new ways to train Neural Networks to learn and predict the solution of the Optimal Power Flow problem, focusing on the optimal control of flexible and controllable units. We propose leveraging the Augmented Lagrangian Method to train a Deep Neural Network by integrating instance-specific Lagrange multipliers and penalties into the loss function. The proposed model is trained on real-world and synthetic data of distribution grids. Numerical tests demonstrate that the suggested method shows significant improvement in runtime compared to a conventional optimizer and can attain near-optimal performance in voltage regulation and loss improvement. Copyright (c) 2024 The Authors.