Physics Informed Neural Networks-Based AC Optimal Power Flow Under High RES Penetration

Creating machine learning techniques to solve AC optimal power flow (ACOPF) problems in power systems requires datasets that are difficult to obtain, do not easily adjust to changing network structures, do not consider worst-case scenarios such as load shedding and renewable energy source (RES) curtailment, and lack the adaptability to handle high levels of RES integration. Moreover, the limited availability or lack of high-quality datasets to train neural networks (NNs) presents a substantial challenge in effectively applying machine learning approaches to address energy distribution issues. This paper proposes a physics-informed neural network (PINN) model that performs complex power system tasks, such as optimal power flow, in less computation time and with higher accuracy, using less data to support high RES penetration. The grid′s physics is represented by ACOPF constraints, using the Karush-Kuhn-Tucker (KKT) formulation. Maximum deviations of the PINN predictions from actual values are evaluated, demonstrating the high performance of the proposed model. Through simulations on the IEEE 6-bus, 9-bus, 24-bus, and 30-bus systems, the proposed PINN model outperforms the standard NN model in computational speed and in predicting generating units' output, load shedding, and RES curtailment under high RES penetration. The most significant deviation in combined generator output prediction is 0.64 MW for the PINN, compared to 0.73 MW for the standard deep NN model. The proposed approach is faster than convex NN (CONN) and Deep Neural Network (DNN), with the highest computational time of 48.3 ms compared to 55.3 ms and 52.6 ms for CONN and NN, respectively. The proposed PINN model demonstrates high accuracy in forecasting generating units' output, RES curtailment, and load shedding, especially under conditions of high RES penetration.