Constraint enforcing for learning the ACOPF using neural networks

This paper explores the application of Deep Neural Networks (DNNs) to solving the Optimal Power Flow (OPF) task. While numerical optimisation techniques have been widely used, their computational complexity limits their real-world applicability. Recent advances in machine learning (ML), in particular the use of Deep Neural Networks (DNNs), offer promising avenues to overcome these limitations. In this effort, we explore the incorporation of the OPF inequality constraints into the architecture of the DNN using activation functions such as clipping, and shifted and scaled sigmoid. Supervised training of the developed models on synthetic datasets demonstrates that the constraints can be enforced at no additional cost, thereby improving prediction accuracy and providing significant speedup compared to the Interior Point Optimiser (IPOPT).