Data-driven AC Optimal Power Flow with Physics-informed Learning and Calibrations

The modern power grid is witnessing a shift in operations from traditional control methods to more advanced operational mechanisms. Due to the nonconvex nature of the Alternating Current Optimal Power Flow (ACOPF) problem and the need for operations with better granularity in the modern smart grid, system operators require a more efficient and reliable ACOPF solver. While data-driven ACOPF methods excel in directly inferring the optimal solution based on power grid demand, achieving both feasibility and optimality remains a challenge due to the NP-hardness of the problem. In this paper, we propose a physics-informed machine learning model and a feasibility calibration algorithm to produce solutions for the ACOPF problem. Notably, the machine learning model produces solutions with a 0.5% and 1.4% optimality gap for IEEE bus 14 and 118 grids, respectively. The feasibility correction algorithm converges for all test scenarios on bus 14 and achieves a 92.2% convergence rate on bus 118.