Data-Driven Modeling of Commercial Off-the-Shelf Photovoltaic Inverters Using Neuromancer

With the high penetration of renewable energy sources into the grid, using inverters with advanced grid support functions has become essential to maintain power quality and reliability. However, it is necessary to have accurate and computationally tractable models of the inverter system to assess the dynamics. Detailed models have been shown to be computationally intractable and require sensitive information from manufacturers to be accurate. Thus, methods that rely on data collection of the inverter system dynamics to analyze the main dynamic behavior of systems under various operating conditions and disturbances are important. In addition, some data-driven models might not be able to achieve high accuracy to assess system dynamics. This paper models a photovoltaic inverter based on the data collected while perturbing the voltage at the point of common coupling and observing the corresponding output current injected into the grid. The model is trained using the PyTorch-based library Neuromancer. A different dataset is used to assess model accuracy and computational time under different Volt-Var support modes. The normalized root mean square error (NRMSE) for each Volt-Var support mode was calculated and compared with other data-driven models in terms of accuracy and computational time. Using the Neuromancer library led to higher accuracy; however, it increased the computational time