Physics-Following Neural Network for Online Dynamic Security Assessment

Data-driven dynamic security assessment (DSA) has emerged as a promising tool for addressing system security challenges posed by the rapid integration of renewable energy resources and power electronic devices. In recent literature, a new concept of physics-informed neural network (PINN), which considers physics described by differential equations, has been deployed for DSA with several benefits. However, existing PINN-based DSA methods face challenges in accurately following dynamic power system physical models due to limitations of algebraic discrepancy, incorrect convergence, and non-convex difficulty during training. To this end, this paper proposes a novel physics-following neural network (PFNN) for DSA by estimating the post-contingency state responses. In particular, a dual-phase training strategy is designed to overcome these specific challenges: (1) a supervised parameter space reduction phase aimed at mitigating non-convex difficulties by initializing the model with empirical loss to enhance optimizability; and (2) a dynamics-guided local learning phase developed to resolve algebraic discrepancies and ensure correct convergence by integrating empirical loss with a physical regularization term derived from dynamic physics models and time-varying algebraic variables. The efficacy of the proposed PFNN is validated through comprehensive case studies conducted on the WSCC 3-machine 9-bus, the New England 10-machine 39-bus, and the IEEE 16-machine 68-bus systems, respectively.