Post-processing seasonal meteorological forecasts with artificial intelligence

Raw meteorological forecasts from global meteorological models are always biased and require post-processing to tailor them to the regional and local climatic features before they can be used for other applications.  However, this might be challenging depending on the features and the meteorological variable considered. This contribution applies and evaluates the use of an artificial intelligence (AI) technique, fuzzy logic (FL), to post-processing meteorological seasonal forecasts, comparing its performance in terms of improved forecasting skills with other post-processing techniques for different forecasting systems and variables. The analysis is applied to the Jucar basins River Basin (Eastern Spain), which are characterized by extreme meteorological events (heavy rains, droughts, heatwaves). For this area, six daily-scale seasonal forecasting systems from the Copernicus Climate Change Service (C3S) and six variables (precipitation; minimum, mean and maximum temperature; solar radiation and wind speed) are considered. ERA5 is used as reference dataset for post-processing, and daily data for the period 1995-2014 is employed to perform the comparison. The evaluation of the performance of AI is done by comparing the skill of AI-based post-processed forecasts with two common post-processing algorithms: linear scaling (LS) and quantile mapping (QM). The algorithms for all three post-processing methods are coded in a Python script. For each system, variable and post-processing alternative, the forecasting skill is measured using the Continuous Range Probability Skill Score (CRPSS). Results show that, with the exception of precipitation, the relative performance of thes methods does not depend on the forecasting system but on the variable considered. FL dominates in maximum and minimum temperature and linear scaling in average temperature, wind speed, and solar radiation. However, LS shows the worst performance in maximum and minimum temperatures, while FL never yields the lowest skill. For precipitation, the ranking between methods depends on the forecasting system. According to the results, FL logic provides robust, skillful post-processing across variables, providing adequate performance for all variables and forecasting systems, while the rest of the methods show a wider spread of performance, from poor to the best. Acknowledgments: This research has been supported by the University Teacher Training (FPU) grant from the Ministry of Universities of Spain (FPU20/0749); the project “INtegrated FORecasting System for Water and the Environment (WATER4CAST)”, funded by the Valencian Government through the Program for the promotion of scientific research, technological development and innovation in the Valencian Community for research groups of excellence, PROMETEO 2021 (ref: PROMETEO/2021/074); and the European Union’s Horizon Europe research and innovation programme under the SOS-WATER project (GA no. 101059264)