A Downscaling Method for Cloud-to-Ground Stroke Density and Its Improvement on Tower-Level Lightning Hazard Risk Identification

Cloud-to-ground (CG) stroke density is key for lightning hazard risk assessment, but conventional grid-cell methods face an inherent trade-off between spatial resolution and statistical confidence. To address this issue, this study takes Fangchenggang City, Guangxi, as the study area and proposes an extreme gradient boosting (XGBoost) regression model integrating high-resolution terrain features to estimate CG stroke density at the hundred-meter scale. The results show that the model produces a 0.001° × 0.001° density distribution that reproduces the spatial pattern of lightning activity. Within the ±30% error interval, the average confidence level of the model output is approximately 44.75%, while that of the grid-cell method is only 0.0163%. The two-dimensional information entropy reaches 10.70 bits, approximately 33.9% higher than that of the grid-cell method. SHAP analysis indicates that the XGBoost model successfully performed smooth interpolation on the raw data to obtain a high-resolution density map showing the effects of topographical factors. Furthermore, a case study of a lightning-induced trip-out event on a 220 kV transmission line illustrates the method’s potential applicability. Compared with kilometer-scale density data, the high-resolution data yield an identified faulty tower that is more consistent with the actual accident location. In summary, we have proposed an effective downscaling method for Cloud-to-Ground stroke density. It is worth noting, however, that independent validation remains limited and that the confidence calculation does not incorporate model uncertainty.