Advanced predictor selection for post-processing of global horizontal irradiance forecasts with stepwise machine learning regression

Read full text Summarise with VIRAT
Authors: Dávid Markovics, Viktória Soós, Martin János Mayer
Cited by: 0
References: 64

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Statistical post-processing utilizing historical data for conditional bias reduction is an essential step in most forecasting workflows. The development of new methods, especially at the advent of advanced machine learning models (ML), is a recurring topic of many post-processing papers. In contrast, little attention is paid to selecting the most informative predictors, even though the accuracy gain achieved through post-processing ultimately depends on the information that the input variables contain regarding the bias of the forecasts. This study compares multiple approaches for ranking 33 potential predictors based on their explanatory power for post-processing the global horizontal irradiance (GHI) forecasts of the European Centre for Medium-Range Weather Forecasts (ECMWF). The proposed stepwise regression (SWR) method is found to be the most reliable feature selection approach, outperforming the Shapley additive explanations method and the feature importance calculations of three tree-based ensemble learning models. Using only the 15 most informative predictors identified by the SWR, a 10.7% reduction in the root mean square error of the already well-calibrated GHI forecasts of the ECMWF has been achieved. Interestingly, the most important predictors are largely different from those commonly used in literature. The SWR process shows great variance over the locations, and proves that a subset of predictors, which provides even greater improvement, exists and can be identified ex post, however, because of year-to-year variability, it cannot be found ex ante. Aggregating the results over the locations improves the generalizability of the feature selection.

Fields of Study & Subjects

Solar Radiation and Photovoltaics Meteorological Phenomena and Simulations Energy Load and Power Forecasting

Publication details

Journal / source
Solar Energy
Volume
311
Pages
114521–114521
Publication date
26 March 2026
Language
EN