Comparison of intraday probabilistic forecasting of solar power using time series models

Abstract

There is a growing interest to estimate the inherent uncertainty of photovoltaic (PV) power forecasts with probability forecasting methods to enable a robust operational planning of energy systems. The aim of this paper is to present a detailed comparison of probabilistic extensions for time series models regarding their practical applicability, as these are still widely used in the field. For this purpose, five probabilistic forecasting methods are analyzed for intraday PV power forecast with respect to their reliability, sharpness and computational complexity. There may be limited data sets available for the initial commissioning in industrial applications. Hence, the analyses are performed using sliding windows of respectively one, three, five and seven weeks of training data. The investigations are carried out with three roof-mounted PV systems from Germany and Austria. An adaptive ARX model is applied as the underlying point forecast, which uses decomposed past PV power measurements and an external weather forecast for the next day as input features. Results suggest that with the methods presented in this paper, already seven days of training data are sufficient to produce significant improvements of up to 45 % in forecasting quality compared to the reference case. Additionally, the quantile regression approaches outperform the bootstrapping approaches slightly in terms of forecast quality and significantly in terms of required computation time.