Transfer Learnable Physics-Informed Neural Network Surrogating Grid-Tied Inverters for Renewable Power System Simulation

The development of accurate and reliable dynamic models of grid-tied inverters is crucial for system-level simulation and the investigation of renewable power systems under multioperating-point (MOP) conditions. However, the limited information available on commercial inverters, due to privacy issues, often renders physics-based modelling methods ineffective. Existing data-driven approaches lack excellent performance when large transients under extreme disturbances are involved, and cannot provide acceptable out-of-domain generalization performance and unified applicability across diverse scenarios. To address the challenges, this article proposes a novel modeling approach, which is driven by a physics-informed neural network (PINN) embedded with differential algebraic equations (DAEs) structure and the customized loss function. This approach enables the creation of high-precision, physically consistent models for inverters without access to internal information. The obtained models can be further encapsulated as surrogate models, which can be seamlessly integrated into a simulation platform as plug-and-play dynamic components, while demonstrating unified applicability across diverse scenarios. Furthermore, cross-extrapolation of surrogate models about different inverters is implemented via transfer learning, maintaining accuracy while significantly conserving computational resources and operational data consumption, which is well-suited for inverters with scarce operational data. The effectiveness and superiority of the proposed approach are validated through various case studies, offering promising solutions for the simulation analysis of a renewable power system incorporating black-box inverters.