Forecasting Convective Gust Potential in the Shanghai Yangtze River Estuary Based on FLBO-CatBoost
-
摘要: 选取2015—2021年逐小时0.25 ° ×0.25 ° ERA5再分析资料和上海长江口区5个沿江地面自动气象站观测资料,运用FLBO-CatBoost集成学习算法,将6种强对流指数作为输入因子,实现对强对流影响下的长江口区强风的潜势分类及概率预报,并利用SHAP方法进行输入因子分析。结果表明:通过加入Multi-Class Focal Loss与贝叶斯优化模块,提高了FLBO-CatBoost综合性能,模型筛选的输入因子物理意义较明确,判断7级强风时POD、CSI、FAR均达到0.70,0.67、0.12,判断8级以上强风时分别达到0.97、0.91、0.07,优于其他五种集成学习模型。运用SHAP方法进行重要性排序可知,在水汽、能量、动力等条件方面,模型能起到优秀的影响要素诊断与筛选功能。同时,将对流云团影响下,长江口区7级以上强风概率预报的最优阈值选定为0.5,之后进一步结合个例验证模型对长江口区的强风的可预报性。整体而言,建立的强风预报模型具有一定的业务应用前景。Abstract: Based on ERA5 reanalysis data and observational data from five automatic meteorological stations in the Yangtze River Estuary from 2015 to 2021, the FLBO-CatBoost is used to classify and predict the probability of convective gusts at level 7 and above in the Yangtze River Estuary. A total of six strong convection indexes are used as input factors and the Shapley additive explanation method is used for factor analysis. The results show that thanks to the incorporation of Multi-Class Focal Loss and Bayesian optimization, the FLBO-CatBoost has performed significantly well. At the same time, the physical meaning of the factors selected by the model is relatively clear. For level 7 convective gusts, the probability of detechion, critical success index, and false alarm ratio values are 0.70, 0.67, and 0.12 respectively. For convective gusts at level 8 and above, they become 0.97, 0.91, and 0.07 respectively. The model outperforms the other five ensemble learning models used in this study. Furthermore, by using the SHAP method for importance ranking, the model demonstrates excellent capacity in diagnosing and selecting influential factors related to moisture, energy, dynamics, and other conditions. In addition, the optimal probability threshold for predicting convective gusts at level 7 and above that are influenced by convective cloud clusters is determined as 0.5. Subsequently, individual cases are examined to further demonstrate the predictability of the model for convective gusts in the region. Overall, the proposed convective gust forecast model proves to be practically useful in the Yangtze River Estuary.
-
Key words:
- convective weather /
- machine learning /
- CatBoost /
- convective gusts /
- potential forecast
-
表 1 模型算法的主要输入因子类型和要素(地面、500 hPa、700 hPa、850 hPa、925 hPa)
要素 主要输入因子 水汽因子 整层水汽含量(TCWV)、露点(Td)、比湿(q)、相对湿度(RH)、温度露点差(T-Td) 动力因子 散度(d)、相对涡度(vo)、垂直速度(w)、位势涡度(PV)、风(u、v) 热力因子 K指数(K)、沙氏指数(SI)、抬升指数(LI)、不稳定能量(CAPE)、总指数(TT)、温度差(T850-500)、温度(T)、对流抑制(CIN) 综合指数 强天气威胁指数(SWEAT)、瑞士雷暴指数(SWISS)、风暴强度指数(SSI)、深对流指数(DCI) 位置信息 经度(lon)、纬度(lat) 高度层因子 0 ℃层高度、-10 ℃层高度 表 2 2015—2021年上海长江口区强风风力占比情况
强风级别 占比 7级(13.9~17.1 m·s-1) 0.661 8级(17.2~20.7 m·s-1) 0.246 9级(20.8~24.4 m·s-1) 0.064 10级以上(≥24.5 m·s-1) 0.029 表 3 检验分类表
实况 预测(无) 预测(有) 实况(无) TN(True negative) FP(False positive) 实况(有) FN(False negative) TP(True positive) -
[1] DOSWELL Ⅲ C A. Severe Convective Storms-An Overview[J]. Severe Convective Storms, 2001, 28(50): l5-25. [2] 齐大鹏, 杨林, 周明飞, 等. 用改进的光流法和卫星TBB资料进行对流临近预报[J]. 热带气象学报, 2023, 39(1): 47-54. [3] 曾琳, 张羽, 李怀宇, 等. 基于多源探测资料的一次广州局地强对流垂直结构分析[J]. 热带气象学报, 2023, 39(3): 348-360. [4] 俞小鼎, 周小刚, 王秀明. 雷暴与强对流临近天气预报技术进展[J]. 气象学报, 2012, 70(3): 331-337. [5] 郑永光, 周康辉, 盛杰, 等. 强对流天气监测预报预警技术进展[J]. 应用气象学报, 2015, 26(6): 641-657. [6] 漆梁波. 高分辨率数值模式在强对流天气预警中的业务应用进展[J]. 气象, 2015, 41(6): 661-673. [7] QX/T 416-2018. 强对流天气等级[S]. 北京: 气象出版社, 2018: 1-2. [8] YU X D, ZHENG Y G. Advances in Severe Convection Research and Operation in China[J], Journal of Meteorological Research, 2020, 34 (1): 1-29. [9] KOLENDOWICZ L, TASZAREK M, CZERNECKI B. Convective and non-convective wind gusts in Poland, 2001-2015, Meteorology Hydrology and Water Management[J]. Research and Operational Applications, 2016, 4(2): 15-21. [10] MOHR S, KUNZ M, RICHTER A, et al. Statistical characteristics of convective wind gusts in Germany[J]. Natural Hazards and Earth System Sciences, 2017, 17(6): 957-969. [11] 郑永光, 宋敏敏. 冷涡影响中国对流性大风与冰雹的分布特征[J]. 热带气象学报, 2021, 37(5/6): 710-720. [12] WAKIMOTO R M. Convectively driven high wind events[M] // DOSWELL Ⅲ C A. Meteor Monogr, Amer Meteor Soc, Boston, 2001, 50: 255-299. [13] STENSRUD D J, XUE M, WICKER L J, et al. Convective-scale warn-on-forecast system: A vision for 2020[J]. Bulletin of the American Meteorological Society, 2009, 90(10): 1 487-1 500. [14] 沈晓玲, 章丽娜, 桑明慧, 等. 浙江夏季弱天气尺度强迫背景下午后局地热对流特征分析[J]. 热带气象学报, 2023, 39(4): 443-452. [15] 朱智慧, 郑运霞, 过霁冰. 上海沿海站点极大风速预报方程的细化和完善[J]. 海洋预报, 2022, 39(1): 32-38. [16] MECIKALSKI J R, WILLIAMS J K, JEWETT C P, et al. ProbabiIistic O-1-h convective initiation nowcasts that combine geostationary satellite observations and numerical weather Drediction model data[J]. J Appl Meteor Climatol, 2015, 54(5): 1 039-1 059. [17] 修媛媛, 韩雷, 冯海磊. 基于机器学习方法的强对流天气识别研究[J]. 电子设计工程, 2016, 24(9): 4 -11. [18] 李文娟, 赵放, 郦敏杰, 等. 基于数值预报和随机森林算法的强对流天气分类预报技术[J]. 气象, 2018, 44(12): 1 555-1 564. [19] SHI X J, CHEN Z, WANG H, et al. Convolutional LSTM network: A machine learning approach for precipitation nowcasting[C]. Advances in Neural Information Processing Systems, 2015: 802-810. [20] TRAN Q K, SONG S. Computer vision in precipitation nowcasting: Applying image quality assessment metrics for training deep neural networks[J]. Atmos, 2019, 10(5): 244. [21] PAN X, LU Y, ZHAO K, et al. Improving nowcasting of convective development by incorporating polarimetric radar variables into a deep learning model[J]. Geophys Res Lett, 2021, 48(21) : 1-10. [22] 李国翠, 刘黎平, 连志鸾, 等. 利用雷达回波三维拼图资料识别雷暴大风统计研究[J]. 气象学报, 2014, 72(1): 168-181. [23] 李海峰. 基于雷达回波的雷暴大风识别算法研究[D]. 黑龙江: 哈尔滨工业大学, 2018. [24] 杨璐, 韩丰, 陈明轩, 等. 基于支持向量机的雷暴大风识别方法[J]. 应用气象学报, 2018, 29(6): 680-689. [25] XIAO H X, WANG Y Q, ZHENG Y, et al. Convective Gusts Nowcasting Based on Radar Reflectivity and a Deep Learning Algorithm[J]. Geoscientific Model Development, 2023, 16(12): 3 611-3 628. [26] 吕晓娜, 牛淑贞, 张一平, 等. 基于概率与权重的雷暴大风潜势客观预报方法研究[J]. 暴雨灾害, 2020, 39(1): 20-29. [27] JOHN T H, TAGHI M K. CatBoost for big data: an interdisciplinary review[J]. Journal of Big Data, 2020, 7(1): 1-45. [28] LIN T Y, GOYAL P, GIRSHICK R, et al. Focal Loss for dense object detection[C]. PROCeedings of the IEEE International Conference on Computer Vision(ICCV), 2017: 2 980-2 988. [29] SHAHRIARI B, SWERSKY K, WANG Z, et al. Taking the human out of the loop: A review of Bayesian optimization[J]. Proceedings of the IEEE, 2016, 104(1): 148-175. [30] 刘建文, 郭虎, 李耀东, 等. 天气分析预报物理量计算基础[M]. 北京: 气象出版社, 2005: 91-141. [31] STEINBACH W R, RICHTER K. Multiple classifiCation and receiver operating characteristic (ROC) analysis[J]. Med Decis Making, 1987, 7(4): 234-237. [32] WOJTUCH A, JANKOWSKI R, PODLEWSKA S. How can shap values help to shape metabolic stability of chemical compounds?[J]. Journal of Cheminformatics, 2021, 13(1): 1-20.