Identification and Correction of Atypical Extreme Heavy Rainfall over the Guangzhou–Foshan Megacity Cluster Based on Key Circulation Factor Clustering

Abstract

This study investigates the relationship between key circulation factors and ensemble forecast uncertainty during an atypical extreme heavy-rainfall event under weak synoptic forcing that affected the Guangzhou–Foshan megacity cluster in the Pearl River Delta (PRD) on 8 September 2022. Here, “atypical” refers to a localized extreme event occurring over the low-elevation urban river network without strong synoptic-scale drivers. Framed as an event-specific retrospective diagnostic analysis, the study used the China Land Multi-source Precipitation Analysis System version 2.1 (CMPAS-V2.1), ERA5 reanalysis, and 3-km (R3) and 9-km (R9) ensemble forecasts from the CMA Tropical Regional Atmospheric Model Ensemble Prediction System (CMA-TRAMS EPS). Spearman rank correlation and Monte Carlo field-significance tests were first applied to identify environmental variables closely associated with hourly precipitation variations, pinpointing the 700-hPa U-wind, 500-hPa V-wind, and 850-hPa V-wind as the most significant circulation predictors. Spatial anomaly fields of these key predictors were then subjected to hierarchical clustering, with clustering robustness evaluated using the cophenetic correlation coefficient (CCC) and bootstrap resampling. Because the clustering structure of the 925-hPa V-wind was comparatively weak, it was excluded from the final member-selection procedure. Finally, circulation-consistent ensemble members were selected using a multi-predictor consensus criterion (retaining 13 R3 members and 7 R9 members), and their 24 h accumulated precipitation was averaged to obtain a circulation-conditioned subset mean. The results show that persistent high temperatures, abundant moisture in the middle and lower troposphere, and a favorable multilayer circulation configuration provided suitable conditions for convective instability accumulation and localized heavy rainfall development. The precipitation forecasts exhibited substantial member-to-member variability. Under a consistent evaluation threshold, the 9-km configuration demonstrated superior overall ensemble-mean spatial skill compared to the 3-km configuration, indicating that increasing horizontal resolution does not necessarily improve forecast skill for weakly forced extreme rainfall. The resulting circulation-conditioned subset means successfully shifted the predicted heavy-rainfall center toward the observed Guangzhou–Foshan region and reduced the overestimated heavy-rainfall magnitudes over northern Guangzhou. Rather than serving as a purely objective score-maximizing post-processing algorithm, this approach extracts physically indicative spatial scenarios from ensemble spread. These findings provide a practical diagnostic framework for conditional forecast correction of localized heavy rainfall under weak synoptic forcing, though validation across multiple independent cases and consistent quantitative verification are necessary to assess its operational generalizability.