Learning ACOPF Without Labels: A Physics-Guided Two-Stage Neural Framework

There has been a growing interest in solving the optimal power flow (OPF) problem using neural networks (NNs) in an unsupervised manner to eliminate the need for labeled OPF instances. This is typically achieved by incorporating physical equations and operational constraints into the loss function. Such approach, however, typically requires both grid states and controls to be included in the output layer even though the core goal of the OPF problem is to determine the optimal controls. Including the full state vector significantly increases the output dimensionality of the neural network, making the architecture unnecessarily large and more difficult to train. To address this limitation, we propose a novel two-stage training framework that decouples the learning of grid physics from the optimization of control actions. This is achieved by introducing two NNs, one learns the power flow equations (PF-NN) and outputs the state of the grid, and another learns to output the controls (Op-NN). Both networks are integrated, and the combined architecture is trained in an unsupervised manner. The resultant Op-NN eventually has only the controls at its output layer enabling reduced size NN while maintaining the unsupervised learning. The case studies show promising results in terms of optimality and in terms of satisfying the physical equations and operational constraints of the grid.