Optimized Forecasting and Verification of Low Visibility for Shanghai Stations Based on Machine Learning
-
摘要: 为提升上海地区大雾等低能见度天气的预报能力，基于机器学习方法建立能见度预报优化模型。研究基于2019年—2023年华东区域业务模式CMA-SH9和天气环境一体化模式WARMS-CMAQ的逐小时预报产品，结合站点能见度等历史气象观测资料，使用LightGBM算法进行建模。为应对低能见度样本稀少导致的观测数据不平衡问题，研究将能见度预报构建为分类任务，并通过数据预清洗以及为不同能见度等级设置差异化权重系数的方式，增强模型对低能见度样本的关注，最终以前两类低能见度事件（vis≤1 km和1 km < vis≤3 km）的TS评分作为核心评估标准。测试集和后续独立运行阶段的评估结果表明，相较于数值模式，机器学习优化模型对能见度的预报评分有显著提升，尤其对于低能见度天气（vis≤1 km）的命中率，从模式的20%左右提升至近60%，TS评分也提升至近0.3。同时，典型案例分析显示LGBM优化模型能够准确地模拟2023年12月份以来的几次大雾过程，且对于大部分过程雾发生和消亡的时间预报与观测更为接近，表明该优化模型相较于传统数值模式能够大幅度提升对大雾过程的模拟能力。Abstract: Based on a machine learning (ML) algorithm, LightGBM, an optimized forecast model of visibility was established to correct the numerical weather prediction (NWP) at Shanghai stations. The model was trained based on the historical station observations (2019-2023) and hourly NWP output data from the numerical weather forecasting (CMA-SH9) and integrated weather-air quality forecasting (WARMS-CMAQ) models. In order to alleviate the problem of extremely unbalanced observational samples and improve the prediction skill for fog and other low-visibility events, visibility was classified into different levels, with the ML task framed as a classification problem. The influence of low-visibility samples was emphasized through data pre-cleaning and differentiated weight coefficients for different grades. Finally, the recall rate, precision, and comprehensive TS scores of the first two grades (≤1 km and 1-3 km) were used as evaluation criteria. The evaluation of the test dataset and subsequent independent operational phase show that the ML-based model significantly improves visibility forecasting skills compared to numerical models. In particular, the hit rate for low visibility events (≤1 km) increased from approximately 20% to nearly 60%, and the TS score improved to 0.3. In addition, the analysis of typical cases since December 2023 shows that the LGBM model performs good in forecasting heavy fogs, with better agreement with observations in terms of fog onset and dissipation. It's indicated that this ML-based model obviously alleviates the serious underprediction of low-visibility events (especially for fog) in the original numerical model, which proves the algorithm's feasibility and superiority.
-
Key words:
- visibility forecasting /
- machine learning /
- fog /
- numerical model
-
图 2 针对能见度分类的前四类等级，两种数值模式和LGBM模型相较于观测在测试集上的五种评分：召回率R（绿色）、精确率P（蓝色）、TS评分（黄色），ETS评分（橙色）
（a）华东区域天气业务模式CMA-SH9；（b）天气-环境一体化模式WARMS-CMAQ；（c）仅考虑模式预报作为输入特征的LGBM模型，其中模式预报综合考虑了天气模式和一体化模式这两种预报产品；（d）在输入特征中加入观测信息后的LGBM模型。
图 3 LightGBM模型的特征重要性（%）分布
分别为每个模式格点所有变量的重要性之和，即格点重要性排序（a），每个模式预报变量的所有格点重要性之和，即模式变量重要性排序（b），及排名前40名的单一特征重要性（c）。其中蓝色代表模式预报变量，其数字下标代表网格；红色代表观测信息，其数字上标代表时间；黄色代表外部的时空信息。
图 4 独立运行阶段EC（蓝色）、WARMS-CMAQ（黄色）以及LGBM优化模型（橙色）能见度站点预报的四种类型的评分：（a）召回率R、（b）精确率P、（c）TS评分、（d）整体的分级相关系数（r），其中，前三类评分分别画出了vis≤1 km和1 < vis≤5 km两个类别的结果，相关系数的计算为了消除高能见度样本对评分的影响将能见度预报数值根据表 3进行分级再计算
图 6 对于2023年12月26日20时—12月30日20时大雾过程，EC（蓝色）、WARMS-CMAQ（绿色）以及LGBM优化模型（红色）对（a）崇明、（b）青浦和（c）南汇的站点能见度预报时间序列与能见度小时观测值的对比，其中灰色代表观测。整个时间序列由每日24~47小时（20：00—次日19：00）的预报时间序列拼接而成。
图 7 同图 6，但时间段为2024年3月28日9时—3月29日20时，站点为（a）崇明、（b）青浦和（c）奉贤。预报序列均为27日晚20时起报的数值模式预报结果和基于此的优化预报。
表 1 机器学习模型输入特征的观测和模式预报产品的变量信息
观测 模式 气象要素 污染物浓度 华东区域业务模式CMA-SH9 天气环境一体化模式 地面预报 高空预报（1 000, 925, 850, 700, 600, 500 hPa） WARMS-CMAQ 能见度VIS $\mathrm{PM}_{10}$ 能见度vis 相对湿度RH 能见度vis_chem 气温TEM $\mathrm{PM}_{2.5}$ 降水APCP $\mathrm{PM}_{10}$ 气压PRES 相对湿度RH CO 相对湿度RH 水平风$U 、V$ $\mathrm{NO}_2$ 水平风$U 、V$ 2 m、地面温度T2M，TMP $\mathrm{O}_3$ 露点温度DPT $\mathrm{PM}_{2.5}$ 反照率REFC $\mathrm{SO}_2$ 
下载: 导出CSV
表 2 LighGBM机器学习模型使用的超参数及其取值
超参数 参数值 学习率 0.05 L1正则化 0.15 L2正则化 0.15 Max_depth 9 Num_leaves 25 Min_data_in_leaf 25
下载: 导出CSV
表 3 能见度等级及其在机器学习模型中的权重系数
能见度vis(km) 能见度等级 权重系数 (0, 1 km] 0 3.9 (1 km, 3 km] 1 1.1 (3 km, 5 km] 2 0.8 (5 km, 10 km] 3 0.45 (10 km, ∞) 4 0.3
下载: 导出CSV
-
[1] Horvath H. Atmospheric visibility[J]. Atmospheric environment, 1981, 15(10-11): 1 785-1 796. [2] Gultepe I, Sharman R, Williams P D, et al. A review of high impact weather for aviation meteorology[J]. Pure and Applied Geophysics, 2019, 176(5): 1 869-1 921. [3] 赵慧霞, 王维国, 李泽椿, 等. 雾对我国交通运输的不利影响及对策[J]. 气象与环境学报, 2010, 26(2): 58-62. [4] 刘喆, 白洁, 张文军. 危险天气导致航空飞行事故的统计分析[J]. 中国民航飞行学院学报, 2022, 33(1): 23-27. [5] 胡伯彦, 冯雷, 陈志豪, 等. 上海地区大面积航班延误预警统计与气象服务质量评估[J]. 中国民航飞行学院学报, 2020, 31(3): 41-45, 58. [6] 上海多车追尾9死43伤或因雾霾[EB/OL]. [2016-11-06], https://www.sohu.com/a/118336150_114731 .[7] 长江口大雾封航致1550艘船积压[EB/OL]. [2018-04-01], https://www.sohu.com/a/226944698_428290 .[8] Bi K, Xie L, Zhang H, et al. Accurate medium-range global weather forecasting with 3D neural networks[J]. Nature (London), 2023, 619(7 970): 533-538. [9] Chen K, Han T, Gong J, et al. Fengwu: Pushing the skillful global medium-range[J]. arXiv: 2304.02948v1, 2023. [10] Lam R, Sanchez-gonzalez A, Willson M, et al. Learning skillful medium-range global weather forecasting[J]. Science, 2023, 382(6 677): 1 416-1 421. [11] Pitchford M, Maim W, Schichtel B, et al. Revised algorithm for estimating light extinction from IMPROVE particle speciation data[J]. Journal of the Air&Waste Management Association, 2007, 57(11): 1 326-1 336. [12] Fast J D, Gustafson W I, Easter R C, et al. Evolution of ozone, particulates, and aerosol direct radiative forcing in the vicinity of Houston using a fully coupled meteorology‐chemistry‐aerosol model[J]. Journal of Geophysical Research: Atmosphere, 2006, 111(D21). [13] Gultepe I, Milbrandt J A. Probabilistic parameterizations of visibility using observations of rain precipitation rate, relative humidity, and visibility[J]. Journal of Applied Meteorology and Climatology, 2010, 49(1): 36-46. [14] Gultepe I, Milbrandt J A, Zhou B. Marine fog: A review on microphysics and visibility prediction[M]//Marine Fog: Challenges and Advancements in Observations, Modeling, and Forecasting, Koračin D, Dorman C E, Cham: Springer International Publishing, 2017, 345-394. [15] Steeneveld G J, Ronda R J, Holtslag A A M. The challenge of forecasting the onset and development of radiation fog using mesoscale atmospheric models[J]. Boundary-Layer Meteorology, 2015, 154(2): 265-289. [16] Yang J, Yu M, Liu Q, et al. A high spatiotemporal resolution framework for urban temperature prediction using IoT data[J]. Computers & Geosciences, 2022, 159: 104991. [17] Prodhan F A, Zhang J, Hasan S S, et al. A review of machine learning methods for drought hazard monitoring and forecasting: Current research trends, challenges, and future research directions[J]. Environmental Modelling & Software, 2022, 149: 105327. [18] Han Y, Zhang G J, Huang X, et al. A moist physics parameterization based on deep learning[J]. Journal of Advances in Modeling Earth Systems, 2020, 12(9)e2020MS002076. [19] Wang A, Xu L, Li Y, et al. Random-forest based adjusting method for wind forecast of WRF model[J]. Computers&Geosciences, 2021, 155: 104842. [20] 徐景峰, 宋林烨, 陈明轩, 等. 冬奥会复杂山地百米尺度10m风速预报的机器学习订正对比试验[J]. 大气科学, 2023, 47(3): 805-824. [21] 方楠, 谢国权, 阮小建, 等. 长短期记忆神经网络(LSTM)模型在低能见度预报中的应用[J]. 气象与环境学报, 2022, 38(5): 34-41. [22] Ortega L C, Otero L D, Solomon M, et al. Deep learning models for visibility forecasting using climatological data[J]. Computers & Geosciences, 2023, 39(2): 992-1004. [23] Duddu V R, Pulugurtha S S, Mane A S, et al. Back-propagation neural network model to predict visibility at a road link-level[J]. Transportation Research Interdisciplinary Perspectives, 2020, 8: 100250. [24] Palvanov A, Cho Y. VisNet: Deep convolutional neural networks for forecasting atmospheric visibility[J]. Sensors, 2019, 19(6): 1343. [25] 祁媛, 樊辉, 张瑞华, 等. 基于ResNet卷积神经网络的机场能见度预测[J]. 自动化应用, 2023, 64(24): 206-209. [26] Bari D. Visibility prediction based on kilometric nwp model outputs using machine-learning regression[C]. IEEE, 2018. [27] Johnson J M, Khoshgoftaar T M. Survey on deep learning with class imbalance[J]. Journal of Big Data, 2019, 6(1): 27. [28] 王勇. 基于多源数据和XGBoost算法的上海市能见度预测模型研究[D]. 上海: 华东师范大学, 2019. [29] Yu Z, Qu Y, Wang Y, et al. Application of machine-learning-based fusion model in visibility forecast: A case study of Shanghai, China[J]. Remote Sensing, 2021, 13(11): 2096. [30] 陈锦鹏, 林辉, 吴雪菲, 等. 基于集成学习的沿海低能见度天气分类预报方法[J]. 热带气象学报, 2023, 39(5): 680-688. [31] Xu Z, Yuhua Y, Baode C, et al. Operational Precipitation Forecast over China using the weather research and forecasting (WRF) model at a gray-zone resolution: Impact of convection parameterization[J]. Weather and Forecasting, 2021, 36(3): 915-928. [32] Xie Y, Xu J, Zhou G, et al. Chemical weather forecasting for eastern China[M]. Air pollution in eastern Asia: An integrated perspective, Bouarar I, Wang X, Brasseur G P, Cham: Springer International Publishing, 2017, 405-416. [33] Brasseur G P, Xie Y, Petersen A K, et al. Ensemble forecasts of air quality in eastern China-Part 1: Model description and implementation of the MarcoPolo-Panda prediction system, version 1[J]. Geoscientific Model Development, 2019, 12(1): 33-67. [34] Xie Y, Wang X, Gao Y, et al. Improving visibility forecasting during haze-fog processes in shanghai and eastern China: The significance of aerosol and hydrometeor extinction[J]. Atmospheric Environment, 2024, 337: 120756. [35] Doswell C A, Davies-jones R, Keller D L. On summary measures of skill in rare event forecasting based on contingency tables[J]. Weather and Forecasting, 2013, 5(4): 576-585. [36] Ke G, Meng Q, Finley T, et al. LightGBM: A highly efficient gradient boosting decision tree[C]. USA: Long Beach, 2017. [37] SchüTz M, SchüTz A, Bendix J, et al. Improving classification‐based nowcasting of radiation fog with machine learning based on filtered and preprocessed temporal data[J]. Quarterly Journal of the Royal Meteorological Society, 2024, 150(759): 577-596. -
点击查看大图
图(7) / 表(3)
计量
- 文章访问数: 2
- HTML全文浏览量: 1
- PDF下载量: 0
- 被引次数: 0
粤公网安备 4401069904700003号