基于低保真物理引导的并行-in-时间神经网络仿真方法用于网络化微电网动态模拟

A Low-fidelity Physics-guided, High-fidelity Neural Simulation (LPHN-Sim) is devised for parallel-in-time dynamic simulation of networked microgrids (NMGs). Three contributions are presented: 1) We design the LPHN-Sim architecture, where a Mori-Zwanzig-based physics solver guides the simulation and a high-fidelity neural network corrects the error; 2) We develop a parallel-in-time reformulation and implementation of LPHN-Sim to circumvent the step-by-step process and enable further acceleration; 3) We train LPHN-Sim via a lightweight physics-informed loss to incorporate dynamic derivatives, which enhances the trainability while maintaining moderate computational cost. Notable features of the method are its ability to generate dynamic trajectories at multiple time points simultaneously and to accurately capture multi-time-scale dynamics. We validate LPHN-Sim in an inverter-dominated NMG under diverse disturbances (e.g., uncertainties, load changes, faults, and cyber interruptions) and demonstrate its accuracy, efficiency, and superiority over purely physics-based or data-driven simulations.