An XGBoost-Based Method for Temperature Forecasting Correction
-
摘要: 以日最高、最低温度以及逐小时整点温度为预报对象,并以ECMWF模式不同起报时间对同一时次的预报项、ECMWF模式前期的预报误差项、前期24小时的实况温度演变项以及用于区分站点(格点)的空间位置信息等共同作为模型的预报因子,采用时空堆叠结合eXtreme Gradient Boosting(XGBoost)方法,对每个预报对象分别构建预报模型,实现了降时间尺度到逐小时的预报以及站点建模到格点预报转化的温度订正预报模型。一年的独立样本试验结果表明:(1) 订正后的日最高和最低温度预报准确率较中央气象台(National Meteorological Center,NMC)温度指导预报产品有明显提高,其中08时和20时起报的预报产品分别提高了28.6%、23.2%和25.5%、16.9%。(2) 逐小时整点温度订正预报中,两种预报产品在每日最高气温出现时段的16时附近均是预报误差比较大的时段;但在绝大多数预报时效的预报中,XGBoost模型显示了比NMC预报产品更好的预报精度。此外,预报个例分析表明,新方案对于较明显的转折天气存在预报滞后性问题。目前,该订正预报方法已实现业务化运行。Abstract: This article focuses on the forecasting of daily maximum and minimum temperatures, as well as hourly temperatures. It utilizes different forecast start times of the ECMWF model for the same time period, early-stage forecast errors of the ECMWF model, the evolution of actual temperatures in the previous 24 hours, and spatial information used to distinguish between stations (grid points) as forecast factors. The approach combines spatiotemporal stacking with the XGBoost method to develop separate forecast models for each temperature variable, achieving a temperature correction forecast model that reduces the time scale to hourly forecasts and transforms from station modeling to grid point forecasting. The results of the independent sample experiment for one year show that (1) the accuracy of the corrected daily maximum and minimum temperature forecast has significantly improved compared to the temperature guidance forecast product (NMC) of the China Meteorological Administration. The forecast products for 08: 00 and 20: 00 have increased by 28.6%, 23.2%, 25.5%, and 16.9%, respectively. (2) In the hourly temperature correction forecast, both forecast products have larger forecast errors around 16:00 when the daily temperature peaks. However, in most forecast periods, the XGBoost model shows better forecast accuracy than the NMC forecast product. Moreover, case analysis of the forecast shows that the new method has a lag problem for more pronounced weather changes. Currently, the temperature forecasting correction method has been implemented for operational use.
-
表 1 XGBoost算法主要参数的默认值
参数 XGBRegressor Max_depth 6 Learning_rate 0.3 Estimator 100 Objective reg: linear Booster gbtree N_jobs 1 Reg_alpha 0 Reg_lambda 1 表 2 三种不同温度预报产品对2021年的预报统计情况
预报对象 检验指标 08 h的预报产品 20 h的预报产品 EC NMC XGBoost EC NMC XGBoost 未来24 h最高温 MAE 3.12 2.08 1.55 3.04 2.25 1.58 ACC 0.35 0.56 0.72 0.37 0.52 0.71 未来24 h最低温 MAE 1.78 1.52 0.93 1.73 1.37 0.95 ACC 0.73 0.73 0.90 0.75 0.77 0.90 -
[1] 孙健, 曹卓, 李恒, 等. 人工智能技术在数值天气预报中的应用[J]. 应用气象学报, 2021, 32(1): 1-11. [2] 谢舜, 孙效功, 张苏平, 等. 基于SVD与机器学习的华南降水预报订正方法[J]. 应用气象学报, 2022, 33(3): 293-304. [3] 韩念霏, 杨璐, 陈明轩, 等. 京津冀站点风温湿要素的机器学习订正方法[J]. 应用气象学报, 2022, 33(4): 489-500. [4] 李倩, 胡邦辉, 王学忠, 等. 基于BP人工神经网络的区域温度多模式集成预报试验[J]. 干旱气象, 2011, 29(2): 231-235. [5] 王焕毅, 谭政华, 杨萌, 等. 三种数值模式气温预报产品的检验及误差订正方法研究[J]. 气象与环境学报, 2018, 34(1): 22-29. [6] 王在文, 陈敏, LUCA D M, 等. 相似集合预报方法在北京区域地面气温和风速预报中的应用[J]. 气象学报, 2019, 77(5): 869-884. [7] 门晓磊, 焦瑞莉, 王鼎, 等. 基于机器学习的华北气温多模式集合预报的订正方法[J]. 气候与环境研究, 2019, 24(1): 116-124. [8] 盛春岩, 范苏丹, 荣艳敏, 等. 几种气温客观预报方法对比及最优集成预报研究[J]. 气象, 2020, 46(10): 1 351-1 361. [9] 吴启树, 韩美, 郭弘, 等. MOS温度预报中最优训练期方案[J]. 应用气象学报, 2016, 27(4): 426-434. [10] 马清, 龚建东, 李莉, 等. 超级集合预报的误差订正与集成研究[J]. 气象, 2008, 399(3): 42-48. [11] 马旭林, 时洋, 和杰, 等. 基于卡尔曼滤波递减平均算法的集合预报综合偏差订正[J]. 气象学报, 2015, 73(5): 952-964. [12] 罗聪, 曾沁, 高亭亭, 等. 精细化逐时滚动温度预报方法及检验[J]. 热带气象学报, 2012, 28(4): 552-556. [13] JEONG S, PARK I, KIM H S, et al. Temperature prediction based on bidirectional long short-term memory and convolutional neural network combining observed and numerical forecast data[J]. Sensors (Basel), 2021, 21(3): 941. [14] RASP S, LERCH S. Neural networks for postprocessing ensemble weather forecasts[J]. Mon Wea Rev, 2018, 146(11): 3 885-3 900. [15] 任萍, 陈明轩, 曹伟华, 等. 基于机器学习的复杂地形下短期数值天气预报误差分析与订正[J]. 气象学报, 2020, 78(6): 1 002-1 020. [16] 尹姗, 李勇, 马杰, 等. 延伸期温度预报误差订正技术初探[J]. 气象, 2020, 46(3): 412-419. [17] CHO D, YOO C, IM J, et al. Comparative assessment of various machine learning‐based bias correction methods for numerical weather prediction model forecasts of extreme air temperatures in Urban Areas[J]. Earth and Space Science, 2020, 7(4): E2019EA000740. [18] JOHNSTONE C, SULUNGU E D. Application of neural network in prediction of temperature: a review[J]. Neural Computing and Applications, 2021, 33(18): 11 487-11 498. [19] FAHIMI NEZHAD E, FALLAH GHALHARI G, BAYATANI F. Forecasting maximum seasonal temperature using artificial neural networks: Tehran case study[J]. Asia-Pacific Journal of Atmospheric Sciences, 2019, 55(2): 145-153. [20] 谭江红, 陈伟亮, 王珊珊. 一种机器学习方法在湖北定时气温预报中的应用试验[J]. 气象科技进展, 2018, 8(5): 46-50. [21] 吴君, 裴洪芹, 石莹, 等. 基于数值预报产品的地面气温BP-MOS预报方法[J]. 气象科学, 2007(4): 430-435. [22] 雷彦森, 蔡晓军, 王文, 等. 遗传算法优化的BP神经网络在地面温度多模式集成预报的应用研究[J]. 气象科学, 2018, 38(6): 806-814. [23] 张亚刚, 王勇, 张肃诏, 等. 基于ECMWF的两种温度客观预报订正方法及评估[J]. 沙漠与绿洲气象, 2022, 16(4): 95-103. [24] 张延彪, 宋林烨, 陈明轩, 等. 基于卷积神经网络的京津冀地区高分辨率格点预报偏差订正试验[J]. 大气科学学报, 2022, 45(6): 850-862. [25] 陈鹤, 蔡荣辉, 陈静静, 等. 基于深度学习方法的气温预报技术应用与评估[J]. 气象, 2022, 48(11): 1 373-1 383. [26] 赵琳娜, 卢姝, 齐丹, 等. 基于全连接神经网络方法的日最高气温预报[J]. 应用气象学报, 2022, 33(3): 257-269. [27] CHEN T, GUESTRIN C. XGBoost: A Scalable Tree Boosting System[C]// the 22nd ACM SIGKDD International Conference. ACM, 2016.