Improving prediction of short-duration heavy rainfall in Guangxi, China during the pre-summer rainy season based on Fengyun-4A lightning frequency and a machine learning algorithm

Weixiang Huang¹Xiaoli Luo^2*Wei Zhang¹Feng Bo³Xiaofei Xia¹Yichen Ying¹Weijun Hu⁴

• ¹Guangxi Power Grid Equipment Monitoring and Diagnosis Engineering Technology Research Center, Electric Power Research Institute of Guangxi Power Grid Co., Ltd., Nanning, China
• ²Guangxi Meteorological Observatory, Nanning, China
• ³Guangxi Key Laboratory of Intelligent Control and operation and Maintenance of Power Equipment, Electric Power Research Institute of Guangxi Power Grid Co., Ltd.,, Nanning, China
• ⁴Fangchenggang Power Bureau of Guangxi Power Grid Co., Ltd,, Fangchenggang, China

In this study, relationships between short-duration heavy rainfall (SDHR) events and lightning activity over Guangxi, China, during the pre-summer rainy season (PSRS) from 2019 to 2023 were investigated via the satellite retrieved IMERG precipitation dataset and Fengyun-4A lightning mapping imager (FY-4A/LMI). The results revealed distinct spatiotemporal variations of the SDHR and lightning activity during the PSRS in Guangxi. SDHR events first occurred in eastern Guangxi in April, expanded westwards by May, and covered the entire region by June. Lightning activity peaked in April, decreased in May, and increased again in June. Both the SDHR and lightning activity exhibited unimodal variations and peaked at nocturnal-to-morning, with the SDHR consistently reaching a maximum intensity at 21:00 UTC—approximately one hour later than the lightning. The mean number of lightning flashes per SDHR event decreased from 8.58 (April) to 6.14 (May) and 6.10 (June). Machine learning experiments revealed that incorporating one-hour antecedent lightning frequency as a predictor in the Random Forest model substantially enhanced the SDHR prediction accuracy. The model demonstrated consistent improvements across all study months, with mean absolute error reductions of 4.42% (April), 6.02% (May), and 4.29% (June). Correspondingly, the coefficient of determination (R2) for SDHR amount increased from 0.29 to 0.35 in April, 0.38 to 0.45 in May, and 0.22 to 0.29 in June. SHapley Additive exPlanation (SHAP) value analysis further confirmed the positive contribution of lightning frequency to rainfall intensity prediction throughout the entire study period and the positive contribution exhibited a monotonically increasing trend when lightning frequency is below the threshold of 15. Lighting frequency can enlarge its influence on the model output via predictor interactions. These results collectively underscore the value of lightning observations as robust predictors for improving short-term heavy rainfall forecasts.