Statistical learning for NWP post-processing: A benchmark for solar irradiance forecasting

The share of solar power in the global and local energy mixes has increased dramatically in the past decade. Consequently, there has been a significant rise in the interest for solar power forecasting, for different time horizons, ranging from few minutes to seasons. For day-ahead forecasts, combination of Numerical Weather Prediction (NWP) models and post-processing algorithms is the most popular approach. Many recent publications have proposed innovative NWP post-processing methods. However, because different works use different datasets, metrics, and even cross-validation methods, it is rarely possible to fairly compare results across several papers. In this work, we propose a rigorous benchmark of several solar NWP post-processing models representative of the literature. For our results to be as general as possible, the comparison is conducted with an open dataset, over 6 years and 7 locations. In addition, we propose a novel benchmarking approach, that focuses on the systematicity of the ranking of forecasting models. Our results show that, when used in combination with proper regularization, large predictor sets are systematically beneficial to NWP post-processing methods. They also demonstrate that more complex algorithms such as neural networks and gradient boosting generally have the lowest mean square error, while support vector regression, a more parsimonious algorithm, performs better in terms of mean absolute error. Lastly, the focus given to ranking systematicity reveals that no model is better in all occasions. This means that researchers should be measured when they conclude to the superiority of a model, in particular when testing data is scarce.

Automated summary

The share of solar power in the global and local energy mixes has significantly increased in the past decade, leading to a rise in interest for solar power forecasting. Numerical Weather Prediction (NWP) models and post-processing algorithms are the most popular methods for day-ahead forecasts. However, comparing results across different studies is challenging due to variations in datasets, metrics, and cross-validation methods. This study proposes a rigorous benchmark of solar NWP post-processing models using an open dataset spanning 6 years and 7 locations. The results demonstrate the systematic benefits of using large predictor sets with proper regularization, as well as the superior performance of more complex algorithms such as neural networks and gradient boosting in terms of mean square error. Support vector regression, a more parsimonious algorithm, performs better in terms of mean absolute error. The study highlights the importance of considering systematic ranking when evaluating forecasting models and emphasizes that no single model is superior in all situations.