Experiment and verification of model clear/rainy forecast correction based on random forest

This study aims to improve clear/rainy forecast accuracy in Anhui Province. Using 3-hour precipitation data (2019-2022) from national/regional stations, CLDAS-V2.0 multi-element grid observations, and numerical model forecasts, we developed a 3-hour smart grid forecast product (RF product) with the Random Forest (RF) algorithm. Seasonal variations and forecast lead times were considered. The RF product was evaluated against frequency matching method (FMM) products, optimal TS score (OTS) products, and raw model forecasts using three metrics: clear/rainy accuracy rate, TS score (≥0.1 mm), and precipitation area overlap ratio. The results are as follows. (1) The RF product showed superior performance. Its 24-hour accuracy (0.855) and TS score (0.611) outperformed all other products. Among models, CMA-SH9 ranked first, followed by CMA-MESO. ECMWF performed worst. FMM and OTS products surpassed ECMWF in accuracy but had higher miss rates. (2) All products achieved better 24-hour forecasts in spring, autumn, and winter. Summer forecasts showed larger errors, but FMM, OTS, and RF products significantly outperformed ECMWF. The RF product maintained consistent advantages across seasons. (3) The RF product achieved higher median and mean overlap ratios in 2022. It best matched actual precipitation areas, particularly reducing position errors in spring and winter. (4) Two case studies demonstrated the RF product’s strengths. Under summer cold shear lines and autumn cyclonic circulation, it achieved >80% and >75% 3-hour accuracy respectively. Its TS scores peaked at 0.55 (summer) and 0.73 (autumn), outperforming other products. The RF product also effectively corrected ECMWF’s 3-hour precipitation location errors.