Learning to Solve the AC-OPF Using Sensitivity-Informed Deep Neural Networks

To shift the computational burden from real-time to offline in delay-critical power systems applications, recent works entertain idea of using a deep neural network (DNN) predict solutions AC optimal flow (AC-OPF) once presented load demands. As topologies may change, training this DNN sample-efficient manner becomes necessity. improve data efficiency, work utilizes fact OPF are not simple labels, but constitute parametric optimization problem. We thus advocate sensitivity-informed (SI-DNN) match only optimizers, also their partial derivatives with respect parameters (loads). It is shown that required Jacobian matrices do exist under mild conditions, and can be readily computed related primal/dual solutions. The proposed SI-DNN compatible broad range solvers, including non-convex quadratically constrained quadratic program (QCQP), its semidefinite (SDP) relaxation, MATPOWER; while seamlessly integrated other learning-to-OPF schemes. Numerical tests on three benchmark corroborate advanced generalization constraint satisfaction capabilities for predicted by an over conventionally trained DNN, especially low-data setups.