Online Dynamic Security Assessment: Using Hybrid Physics-Guided Deep Learning Models

Two main tasks of online dynamic security assessment (DSA) are real-time state monitoring and postfault transient trajectory prediction. For the first task, the model-based methods are too computationally intensive for real-time state estimation/prediction. As for the second task, the existing deep learning-based methods always require massive training data and fail to provide accurate long-term prediction given the limited fault-on data. To solve the abovementioned problems, a novel online DSA based on physics-guided deep learning models is developed. First, a deep Koopman operator (DKO) is developed to model the system dynamics. Thanks to the linear prediction capability of the DKO, it can be used not only to predict future states from previous states but also to develop a deep Koopman Kalman filter to estimate states from measurements. Then, a physics-informed neural network (PINN) is developed. By integrating the swing equation into the model training, the trained PINN can predict the postfault transient trajectory directly from the fault-on data without waiting for the postfault initial state trajectory or external data. Comparison experiments are conducted on the IEEE 39 bus system and the IEEE 118 bus system to verify the effectiveness and efficiency of the developed approach.