Recurrent Neural Networks for Solving Photovoltaic System Dynamics

This paper addresses computational challenges in dynamic simulation of distribution grids with high penetration of photovoltaic (PV) systems. Electromagnetic transient (EMT) and phasor-based PV models are computationally intractable for analyzing PV-rich large-scale distribution grids. Hence, this study explores Recurrent Neural Network (RNN)-based dynamic equivalent models (DEMs), specifically, Long Short-Term Memory (LSTM) and Non-linear Auto Regressive with eXogenous input (NARX), for accurate and computationally efficient representation of PV system dynamics. The online efficacy of these DEMs is compared using the IEEE 13-node test feeder with 32 PV units, constituting 25% PV-based generation. Both LSTM and NARX based DEMs demonstrate similar learning capabilities and offline performances, yielding testing efficiencies of 99.64% and 99.95%, respectively. During online testing, LSTM and NARX based DEMs exhibit compatible performances with Mean-Absolute-Errors (MAEs) of 0.81 and 0.17, respectively, while achieving accelerated performances compared to base model. NARX-based DEM outperforms LSTM-based counterpart with a remarkable 40 times speedup.