| 摘 要: 针对传统的雷云识别与追踪聚类算法大多依赖人工设定参数且在雷电临近预测中雷云的运动通常被简化为 线 性 轨 迹 的 问 题,提 出 了 多 密 度 自 适 应 DBSCAN(Density-BasedSpatialClusteringofApplicationswithNoise)算法,算法以动态调整参数来适应数据集中不同簇的密度差异。在聚类分析结果的基础上,结合极限梯度提升(XGBoost)学习历史雷云质心的运动规律对雷云进行预测,并利用反距离权重插值法和正态分布计算雷云区域。结合实际发生的雷暴活动进行验证分析。实验结果表明:结合多密度自适应 DBSCAN的 XGBoost算法相较于传统的线性外推和最小二乘法预测雷云质心的效果有明显的提升,平均误差在20km 以下,且在15min内预测落雷点,对比实际落雷点的虚警率在25%以下、漏警率在35%以下,为江西省的雷电防范提供了参考。 |
| 关键词: 雷电临近预警 闪电定位 DBSCAN XGBoost |
|
中图分类号: TP391
文献标识码: A
|
| 基金项目: 国家自然科学基金项目资助(61672293) |
|
| Lightning Proximity Forecasting in Jiangxi Province Based on XGBoost Algorithm |
|
HUANG Qiuyu1, CHEN Xinyu2, ZHANG Jianwei1, XU He1
|
(1.School of Mathematics and Statistics, Nanjing University of Information Science & Technology, Nanjing 210044, China;
2.Jiujiang Meteorological Bureau, Jiujiang 332000, China)
927874145@qq.com; 1553722802@qq.com; zhangjw@nuist.edu.cn;; 529110275@qq.com
|
| Abstract: Addressing the issue that traditional thundercloud identification and tracking clustering algorithms largely rely on manually set parameters and often simplify thundercloud movement to linear trajectories in lightning proximity prediction, this study proposes a mult-i density adaptive DBSCAN (Density-Based Spatial Clustering of Applications with Noise)algorithm to dynamically adjust parameters to accommodate density differences among clusters in the dataset. Based on the clustering analysis results, the Extreme Gradient Boosting (XGBoost) algorithm is used to learn the movement patterns of historical thundercloud centers for prediction purposes. The thundercloud area is calculated using the inverse distance weighting interpolation method and a normal distribution. The model is validated through analysis of actual thunderstorm activity. Experimental results indicate that the XGBoost algorithm combined with the mult-i density adaptive DBSCAN method significantly improves the accuracy of predicting thundercloud centers compared to traditional linear extrapolation and least squares methods. The average error is below 20 km, and within 15 minutes, the false alarm rate for predicting lightning strike points compared to actual strike points is below 25% , and the missed alarm rate is below 35% . This provides a reference for lightning protection in Jiangxi Province. |
| Keywords: lightning proximity lightning positioning data DBSCAN XGBoost |
