A study on fine scale precipitation nowcasting in Beijing-Tianjin-Hebei region based on deep learning
-
摘要: 精细尺度降水的临近预报对于提升现代城市内涝和山洪地质灾害预警能力具有重要意义。深度学习作为一种新兴方法，在挖掘数据内部特征及物理规律方面更具优势，近年来在天气雷达图像领域的应用已初见成效。为进一步提升精细尺度降水的临近预报能力，基于深度学习网络模型RainNet，研究建立了两种滚动预报方式，开展了京津冀地区1 km分辨率精细尺度降水滚动式临近预报试验和对比分析。 试验结果表明：与传统基于交叉相关的外推预报相比，深度学习网络模型RainNet总体可以明显改进降水1 h临近预报的绝对误差和相关系数；两个RainNet相结合的滚动预报方式对1.04 mm/（10 min）及以下阈值降水，在10—50 min预报性能一致优于传统的交叉相关外推预报。深度学习模型对降水消亡过程的时、空演变趋势刻画更好，尤其更适用于降水消亡过程的临近预报。采用两个RainNet模型相结合的滚动式预报方式优于单一模型滚动预报方式。Abstract: Precipitation nowcasting on fine scale is of great significance to improve the ability of early warning of flood and waterlogging disasters in modern cities. As a new method, deep learning has more advantages in mining the internal characteristics and physical laws of data. In recent years, the application of deep learning in the field of meteorological radar image has achieved preliminary results. In order to improve the effectiveness of nowcasting on fine scale, a deep convolutional neural network-RainNet is used to propose two ways of rolling approach for precipitation nowcasting. Experiments and comparative analysis are carried out in Beijing-Tianjin-Hebei region on 1 km resolution. Compared with the traditional extrapolation based on Tracking Radar Echoes by Correlation (TREC), the results show that the mean absolute error and correlation coefficient of 1 h nowcasting can be improved. The prediction in 10—50 min in thresholds of 1.04 mm/(10 min) and below is better than that of traditional prediction. Temporal and spatial evolution of precipitation extinction process is better described by deep learning compared with that by traditional extrapolation. The rolling approach with two RainNet models combined outperforms one single model in precipitation nowcasting.
-
Key words:
- Precipitation /
- Nowcasting /
- Deep learning /
- TREC /
- Extrapolation
-
图 1 京津冀地区地形海拔高度 （单位：m） 及8部天气雷达位置 （“+”表示）
Figure 1. Surface elevation （unit：m） around the study area in contiguous North China and the eight radar sites of the operational China Next Generation Weather Radar network are displayed （+ symbols）
图 2 研究区域范围 （虚线矩形框为原始降水分析场的覆盖范围581 km×511 km，实线矩形为裁剪后研究范围576 km×496 km）
Figure 2. Study domain （the dotted rectangle is the domain of 581 km×511 km for original precipitation analysis，the rectangle within the solid line is the final study domain of 576 km×496 km）
图 4 基于深度学习网络模型的滚动式临近预报原理示意
Figure 4. Schematic diagram of rolling nowcast based on deep learning network model
图 5 交叉相关方法示意 （图a、b分别代表t、t+Δt时刻降水场；黑色矩形框代表求交叉相关的像素点区域，r代表搜索半径，黑色矢量为TREC矢量，所指向终点代表所匹配的具有最大相关系数的矩形中心）
Figure 5. Schematic diagram showing the computation of the TREC （Fig. 5a and Fig. 5b represent precipitation analysis at time t and t+Δt respectively，r is search radius，the black vector is the TREC vector，the end point of which is the centre of the array at time t+Δt with maximum correlation coefficient）
图 6 2020年6—9月的绝对误差空间分布 （单位：mm/（10 min）；a. 基于交叉相关的外推预报 （TREC） 绝对误差，b. 两个深度学习模型相结合的滚动预报 （RainNet） 绝对误差；下标1—6分别表示预报时效为10、20、30、40、50、60 min）
Figure 6. Spatial distributions of mean absolute error （MAE） during June to Septemper 2020 （unit：mm/（10 min）； a. TREC predictions，b. two rolling RainNet predictions；1—6 represents the forecast time in 10，20，30，40，50 and 60 min，respectively）
图 7 2020年6—9月京津冀地区的相关系数空间分布 （a. 基于交叉相关的外推预报 （TREC） 相关系数，b. 两个深度学习模型相结合滚动预报 （RainNet） 的相关系数；下标1—6分别表示预报时效为10、20、30、40、50、60 min）
Figure 7. Spatial distributions of correlation coefficient during June to September 2020 （a. TREC predictions，b. RainNet predictions；1—6 represents the forecast time in 10，20，30，40，50 and 60 min，respectively）
图 8 绝对误差的时间序列箱线图 （黑色方框的上、下边界为75%、25%分位数，上须和下须分别是90% 和10%分位数，黑盒子中的白线和白三角分别是中位数和均值；a—f. 分别是预报时效为10、20、30、40、50、60 min的结果）
Figure 8. Box plots of mean absolute error （the upper and lower of the black box are the quantiles of 75% and 25%，the upper and lower whiskers are 90% and 10% quantiles，the white line and triangle in the black box show the median and mean values，respectively；a—f. the forecast time is 10，20，30，40，50 and 60 min，respectively）
图 9 相关系数的时间序列箱线图 （黑色方框的上、下边界为75%、25%分位数，上须和下须分别是90% 和10%分位数，黑盒子中的白线和白三角分别是中位数和均值；a—f. 分别是预报时效为10、20、30、40、50、60 min的结果）
Figure 9. Box plots of correlation coefficient （the upper and lower of the black box are the quantiles of 75% and 25%，the upper and lower whiskers are 90% and 10% quantiles，the white line and triangle in the black box show the median and mean values，respectively；a—f. the forecast time is 10，20，30，40，50 and 60 min，respectively）
图 10 2020年6—9月降水临近预报的相关系数 （a） 和绝对误差 （b） 曲线
Figure 10. Mean absolute errors （a） and correlation coefficients （b） during June to September 2020
图 11 2020年6—9月不同降水阈值 （a—d） 的CSI评分曲线
Figure 11. Critical success indexes （CSI） for four different intensity thresholds （a—d） during June to September 2020
图 12 2020年7月31日11时20分起报10—60 min （a1—a6、b1—b6） 降水临近预报结果及降水实况（c1—c6） （a. 交叉相关外推预报，b. 深度学习RainNet预报，c. 相应时间的降水实况）
Figure 12. Predictions of TREC （a1—a6） and RainNet （b1—b6） as well as corresponding observations （c1—c6） at 11：20 UTC 31 July 2020 （1—6 represents the forecast time in 10，20，30，40，50 and 60 min，respectively）
图 13 2020年8月1日13时50分起报10—60 min （a1—a6、b1—b6） 降水临近预报结果及降水实况 （c1—c6）（a. 交叉相关外推预报，b. 深度学习RainNet预报，c. 相应时间的降水实况）
Figure 13. Predictions of TREC （a1—a6） and RainNet （b1—b6） as well as corresponding observations （c1—c6） at 13：50 UTC 1 August 2020 （1—6 represents the forecast time in 10，20，30，40，50 and 60 min，respectively）
图 14 2020年8月15日03时00分起报10—60 min （a1—a6、b1—b6） 降水临近预报结果及降水实况 （c1—c6）（a. 交叉相关外推预报，b. 深度学习RainNet预报，c. 相应时间的降水实况）
Figure 14. Predictions of TREC （a1—a6） and RainNet （b1—b6） as well as corresponding observations （c1—c6） at 03：00 UTC 15 August 2020 （1—6 represents the forecast time in 10，20，30，40，50 and 60 min，respectively）
表 1 模型数据集信息
Table 1. Main parameters of the datasets
年份 训练集 测试集 2017 2018 2019 2020 样本数 2852 3199 2957 3266 总计 9008 3266 
下载: 导出CSV
-
[1] 曹伟华,陈明轩,高峰等. 2019. 雷暴区域追踪矢量与雷暴单体追踪矢量融合临近预报研究. 气象学报,77(6):1015-1027 doi: 10.11676/qxxb2019.064Cao W H,Chen M X,Gao F,et al. 2019. A vector blending study based on object-based tracking vectors and cross correlation tracking vectors. Acta Meteor Sinica,77(6):1015-1027 (in Chinese) doi: 10.11676/qxxb2019.064 [2] 陈明轩,高峰,孔荣等. 2010. 自动临近预报系统及其在北京奥运期间的应用. 应用气象学报,21(4):395-404 doi: 10.3969/j.issn.1001-7313.2010.04.002Chen M X,Gao F,Kong R,et al. 2010. Introduction of auto-nowcasting system for convective storm and its performance in Beijing Olympics meteorological service. J Appl Meteor Sci,21(4):395-404 (in Chinese) doi: 10.3969/j.issn.1001-7313.2010.04.002 [3] 程丛兰,陈敏,陈明轩等. 2019. 临近预报的两种高时空分辨率定量降水预报融合算法的对比试验. 气象学报,77(4):701-714 doi: 10.11676/qxxb2019.017Cheng C L,Chen M,Chen M X,et al. 2019. Comparative experiments on two high spatiotemporal resolution blending algorithms for quantitative precipitation nowcasting. Acta Meteor Sinica,77(4):701-714 (in Chinese) doi: 10.11676/qxxb2019.017 [4] 郭瀚阳,陈明轩,韩雷等. 2019. 基于深度学习的强对流高分辨率临近预报试验. 气象学报,77(4):715-727 doi: 10.11676/qxxb2019.036Guo H Y,Chen M X,Han L,et al. 2019. High resolution nowcasting experiment of severe convections based on deep learning. Acta Meteor Sinica,77(4):715-727 (in Chinese) doi: 10.11676/qxxb2019.036 [5] 韩丰,龙明盛,李月安等. 2019. 循环神经网络在雷达临近预报中的应用. 应用气象学报,30(1):61-69 doi: 10.11898/1001-7313.20190106Han F,Long M S,Li Y A,et al. 2019. The application of recurrent neural network to nowcasting. J Appl Meteor Sci,30(1):61-69 (in Chinese) doi: 10.11898/1001-7313.20190106 [6] 孙健,曹卓,李恒等. 2021. 人工智能技术在数值天气预报中的应用. 应用气象学报,32(1):1-11 doi: 10.11898/1001-7313.20210101Sun J,Cao Z,Li H,et al. 2021. Application of artificial intelligence technology to numerical weather prediction. J Appl Meteor Sci,32(1):1-11 (in Chinese) doi: 10.11898/1001-7313.20210101 [7] 王改利,刘黎平,阮征. 2007. 多普勒雷达资料在暴雨临近预报中的应用. 应用气象学报,18(3):388-395 doi: 10.3969/j.issn.1001-7313.2007.03.016Wang G L,Liu L P,Ruan Z. 2007. Application of doppler radar data to nowcasting of heavy rainfall. J Appl Meteor Sci,18(3):388-395 (in Chinese) doi: 10.3969/j.issn.1001-7313.2007.03.016 [8] Ayzel G,Heistermann M,Sorokin A,et al. 2019. All convolutional neural networks for radar-based precipitation nowcasting. Procedia Comput Sci,150:186-192 doi: 10.1016/j.procs.2019.02.036 [9] Ayzel G,Scheffer T,Heistermann M. 2020. RainNet v1.0:A convolutional neural network for radar-based precipitation nowcasting. Geosci Model Dev,13(6):2631-2644 doi: 10.5194/gmd-13-2631-2020 [10] Chen L,Cao Y,Ma L M,et al. 2020. A deep learning-based methodology for precipitation nowcasting with radar. Earth Space Sci,7(2):e2019EA000812 [11] Chen M,Bica B,Tüchler L,et al. 2017. Statistically extrapolated nowcasting of summertime precipitation over the Eastern Alps. Adv Atmos Sci,34(7):925-938 doi: 10.1007/s00376-017-6185-4 [12] Dixon M,Wiener G. 1993. TITAN:Thunderstorm identification,tracking,analysis,and nowcasting:A radar-based methodology. J Atmos Oceanic Technol,10(6):785-797 doi: 10.1175/1520-0426(1993)010<0785:TTITAA>2.0.CO;2 [13] Germann U,Zawadzki I. 2002. Scale-dependence of the predictability of precipitation from continental radar images. Part I:Description of the methodology. Mon Wea Rev,130(12):2859-2873 doi: 10.1175/1520-0493(2002)130<2859:SDOTPO>2.0.CO;2 [14] Haiden T,Kann A,Wittmann C,et al. 2011. The integrated nowcasting through comprehensive analysis (INCA) system and its validation over the eastern Alpine region. Wea Forecasting,26(2):166-183 doi: 10.1175/2010WAF2222451.1 [15] Han L,Sun J Z,Zhang W,et al. 2017. A machine learning nowcasting method based on real-time reanalysis data. J Geophys Res,122(7):4038-4051 doi: 10.1002/2016JD025783 [16] Johnson J T,MacKeen P L,Witt A,et al. 1998. The storm cell identification and tracking algorithm:An enhanced WSR-88D algorithm. Wea Forecast,13(2):263-276 doi: 10.1175/1520-0434(1998)013<0263:TSCIAT>2.0.CO;2 [17] Kingma D P, Ba J. 2015. Adam: A method for stochastic optimiza- tion∥Proceedings of the 3rd International Conference on Learning Representations. San Diego: ICLR [18] Li L,Schmid W,Joss J. 1995. Nowcasting of motion and growth of precipitation with radar over a complex orography. J Appl Meteorol Climatol,34(6):1286-1300 doi: 10.1175/1520-0450(1995)034<1286:NOMAGO>2.0.CO;2 [19] Mueller C,Saxen T,Roberts R,et al. 2003. NCAR auto-nowcast system. Wea Forecast,18(4):545-561 doi: 10.1175/1520-0434(2003)018<0545:NAS>2.0.CO;2 [20] Rinehart R E,Garvey E T. 1978. Three-dimensional storm motion detection by conventional weather radar. Nature,273(5660):287-289 doi: 10.1038/273287a0 [21] Shi X J,Chen Z R,Wang H,et al. 2015. Convolutional LSTM network:A machine learning approach for precipitation nowcasting∥Proceedings of the 28th International Conference on Neural Information Processing Systems. Montreal:MIT Press,802-810 [22] Sokol Z,Kitzmiller D,Pesice P,et al. 2013. Comparison of precipitation nowcasting by extrapolation and statistical-advection methods. Atmos Res,123:17-30 doi: 10.1016/j.atmosres.2012.10.013 [23] Staniforth A,Côté J. 1991. Semi-Lagrangian integration schemes for atmospheric models:A review. Mon Wea Rev,119(9):2206-2223 doi: 10.1175/1520-0493(1991)119<2206:SLISFA>2.0.CO;2 [24] Sun J Z,Xue M,Wilson J W,et al. 2014. Use of NWP for nowcasting convective precipitation:Recent progress and challenges. Bull Amer Meteor Soc,95(3):409-426 doi: 10.1175/BAMS-D-11-00263.1 [25] Wilks D S. 2011. Statistical Methods in the Atmospheric Sciences. 3rd ed. Boston:Academic Press,309-310 [26] Wilson J W,Crook N A,Mueller C K,et al. 1998. Nowcasting thunderstorms:A status report. Bull Amer Meteor Soc,79(10):2079-2100 doi: 10.1175/1520-0477(1998)079<2079:NTASR>2.0.CO;2 -
下载:
点击查看大图
计量
- 文章访问数: 516
- HTML全文浏览量: 99
- PDF下载量: 287
- 被引次数: 0
