Time-Domain Simulation based on Physics-Informed DeepONet

Online security assessment requires fast and reliable time-domain prediction of power-system dynamic responses. However, numerical-integration-based methods are often excessively time-consuming, while existing data-driven approaches usually have limited robustness and physical consistency. This paper proposes a time-domain simulation method based on Physics-informed Deep Operator Network (PI-DeepONet). The PI-DeepONet is designed to incorporate ordinary differential equations (ODE) into the loss function as a physical regularization term, thus enhancing physical consistency and interpretability. Given initial operating conditions, the PI-DeepONet approximates the trajectory of generator dynamics. The estimated generator dynamics, together with algebraic equations, are used to iteratively perform multi-step simulations. In the single-machine case, PI-DeepONet reduces the computation time by 93.86% compared with the Runge-Kutta method while keeping the trajectory errors below 1%, demonstrating that the proposed method outperforms numerical integration in computational efficiency. Furthermore, the PI-DeepONet exhibits higher accuracy and better generalization capability than PINN methods. A preliminary three-generator case further shows that the proposed method can be embedded into multi-machine DAE simulation. These results indicate that the proposed method has practical potential for online time-domain simulation and dynamic security assessment.