Equivalent Model of Photovoltaic System Dynamics Using Neural Network

Dynamic simulation of distribution grids with high penetration of photovoltaic (PV) systems is computationally demanding. The existing PV models, including electromagnetic transient (EMT)-based detailed models and phasor-based approaches, pose computational challenges when applied to system-level studies involving numerous PV units. As power distribution grids integrate more dynamic units (e.g., PV inverters), there is an immediate need for computationally tractable dynamic PV models with grid support functionalities. In this work, we sought to develop Recurrent Neural Network (RNN)-based dynamic equivalent model (DEM) for smart PV systems using Nonlinear Autoregressive with Exogenous input (NARX) type model. The efficacy of the proposed approach is demonstrated using the IEEE 13-node test feeder, with 32 PV units representing 25% PV generation. Case studies at 25% PV penetration demonstrate that the proposed NARX-RNN-based DEM achieves comparable accuracy with mean-squared-errors (MSEs) of 0.0215 kW for active power and 0.0076 kVAr for reactive power responses. Notably, the NARX-RNN model surpasses the phasor-based approach by over 80 times and the EMT model by over 4,000 times in terms of simulation time.