Statistical postprocessing of ensemble global radiation forecasts with penalized quantile regression

Abstract

Nowadays, ensemble-based numerical weather forecasts provide probabilistic guidance to actors in the renewable energy sector. Ensemble forecasts can however suffer from statistical inconsistencies that affect the forecast reliability. Statistical post-processing techniques address this issue using learning algorithms based on past data. In this study, it is shown that quantile regression is a suitable method for the post-processing of ensemble global radiation forecasts. In a basic approach, conditional quantiles are estimated using the first guess quantile forecasts and a solar geometry variable as predictors. In a more complex approach, adequate meteorological predictors are selected among a pool of ensemble model outputs by means of a regularization scheme. The so-called penalized quantile regression and the basic quantile regression approaches, respectively, are applied to hourly averaged global radiation forecasts of the high-resolution ensemble prediction system COSMO-DE-EPS. Both calibration setups provide reliable probabilistic forecasts at all investigated probability levels, which improves considerably the ensemble forecast skill. Moreover, verification results demonstrate that including rigorously selected predictors in the regression scheme increases the ensemble forecast sharpness and thereby the value of the probabilistic guidance.

Keywords

• calibration
• ensemble forecast
• global radiation
• penalized regression
• quantile regression