Integrating Multi-Neural DAE Experts into Power System Dynamic Component Modeling for EMTP-Type Simulation

Traditional data-driven modeling methods encounter significant accuracy limitations when used to model the electromagnetic transients (EMT) of unidentified components in power systems based on limited measurement data. Moreover, their restricted generalization capabilities and compatibility issues further impede broad adoption. This paper proposes a novel multi-gate mixture-of-experts based differential-algebraic equation (MMOE-DAE) network to model components with unknown high-order dynamics. First, an MMOE-DAE network is established to model the relationships among the phases of the three-phase component by modulating multiple gated neural DAEs. Then, an automatic order upgrading module for high-order dynamic component modeling is proposed, and the identification of DAE order is facilitated by a fully gated network. Last, a model local parameter flexible calibration strategy based on fine-tuning is designed to improve the accuracy and stability of the MMOE-DAE training. Based on the commercial EMT simulation platform CloudPSS, dynamic equivalent models for different power components are constructed, and the accuracy and flexibility advantages of MMOE-DAE modeling are assessed through co-simulations.