Comparing support vector regression for PV power forecasting to a physical modeling approach using measurement, numerical weather prediction, and cloud motion data

Abstract The growth of installed photovoltaic (PV) power capacity in recent years has emerged an increasing interest in high quality forecasts. The most common ways to predict PV power output are either applying statistical approaches to PV measurements or calculating future outputs of a PV module with known specification applying a PV simulation model to irradiance forecasts. In this work, we compare these two concepts to a statistical learning model, i.e., support vector regression (SVR), that is applied on a large dataset of PV power measurements, numerical weather predictions, and satellite-based cloud motion vector forecasts. To achieve a high forecast accuracy with SVR, we first perform an extensive parameter optimization on a subset of all available PV systems for pre-selected days. We limit the input features of the SVR to those of the other models to increase comparability between the different approaches. Despite these limitations, the SVR shows promising results, especially in comparison with the physical approaches without any statistical improvements. A SVR forecasting model that combines all input features is able to generate predictions with a similar accuracy as statistically enhanced predictions of a PV simulation model.