Comparison of a recurrent neural network PV system model with a traditional component-based PV system model

Photovoltaic (PV) system modeling is used throughout the photovoltaic industry for the prediction of PV system output under a given set of weather conditions. PV system modeling has a wide range of uses including: pre-purchase comparisons of PV system components, system health monitoring, and estimation of payback (return on investment) times. In order to adequately model a PV system, the system must be characterized to establish the relationship between given weather inputs (e.g., irradiance, spectrum, temperature) and desired system outputs (e.g., AC power, module temperature). Traditional approaches to system characterization involve characterizing and modeling each component in a PV system and forming a system model by successively using component models. This paper compares a traditional modeling approach using the Sandia Photovoltaic Array Performance Model [1] to a new method of characterization using a recurrent neural network (RNN). The Sandia model predicts system performance from given weather data and individual component characterizations using a defined set of equations, while the RNN “learns” the input/output relationships by training on concurrent weather and performance data. The comparison of a traditional modeling technique and the new RNN method serves to validate the accuracy of the new method in comparison to a widely accepted modeling technique. Modeling using an RNN may be advantageous when component models are not available for the components in a PV system, when the components of a PV system are unknown to the modeler, or when system components are installed or altered in such a fashion that their model parameters are no longer applicable.