| 作者单位 |
1. Inner Mongolia Meteorological Observatory, Hohhot 010051, China; 2. Ulanqab Meteorological Bureau, Ulanqab 012000, China; 3. Siziwang Banner Meteorological Bureau, Inner Mongolia Autonomous Region, Siziwang Banner 011899, China |
| 摘要 |
Lightning activity is characterized by abrupt occurrence, strong spatial heterogeneity, and highly variable intensity, and traditional empirical methods are difficult to meet the needs of refined lightning warning. In this paper, the Ulanqab region was selected as the study area.VLF/LF lightning location data and ERA5 reanalysis data collected during the summers of 2020-2022 were used, and a lightning intensity level prediction model was constructed. Based on the environmental conditions of lightning occurrence, nine key predictors were selected from 25 environmental factors, including chimney index, total totals index, K index, 0-6 km of vertical wind shear, deep convection index, uplift condensation height temperature, dew point temperature, 700 hPa of pseudo equivalent temperature, and precipitable water vapor. Four machine learning methods, namely gradient boosting, random forest, neural network, and logistic regression, were used to establish a five-level lightning intensity classification model for weak, moderate, strong, very strong, and extreme lightning. The results showed that the gradient boosting model had the highest comprehensive score of 0.173 7, and performed the best in terms of average CSI, overall ACC, and comprehensive prediction ability. The DeLong test results indicated that the overall performance of gradient boosting was significantly better than other models (P<0.05). Using the typical lightning process during August 18-20, 2024, a 72-hour historical simulation and EC business prediction verification were conducted. The results showed that the model can better reflect the changes in lightning activity intensity and spatial evolution patterns, and has the business forecasting application capability. |
