设为首页
加入收藏
The Application of Recurrent Neural Network to Nowcasting
-
摘要: 该文将循环神经网络（recurrent neural network，RNN）应用于雷达临近预报。使用预测循环神经网络（predictive RNN）架构，利用雷达历史组合反射率因子建模，给出雷达组合反射率因子未来1 h的预报结果。预测循环神经网络的核心是在长短时记忆单元（long short-term memory，LSTM）中增加时空记忆模块，能够提取雷达回波不同尺度的空间特征，配合循环神经网络架构，可以有效解决反射率因子预测问题。北京大兴雷达和广州雷达长时间序列的独立检验结果和2个强对流天气个例检验结果表明：该方法和传统的基于交叉相关法的1 h雷达外推临近预报相比，在20 dBZ和30 dBZ检验项目内，临界成功指数（CSI）可以提升0.15~0.30，命中率（POD）提高0.15~0.25，虚警率（FAR）降低0.15~0.20，该方法对反射率因子强度变化有一定预报能力。Abstract: Radar extrapolation is an important means in nowcasting. The radar extrapolation methods widely used in China include COTREC and Optical Flow, by which two consecutive echoes are used to diagnose the advection velocity within rain analyses, involving the solution of Lagrangian persistence equation. A new method RNN(recurrent neural network) is applied in nowcasting. Using PredRNN(predictive RNN), by modeling historical radar data, the prediction of radar echo in the next hour is given. PredRNN consists of ST-LSTM unit, which is an improvement of LSTM. One advantage of using PredRNN is the operation of the state accumulation and the hidden layer output is replaced by convolution. Therefore, the neurons not only can get timing relationships, but also extract spatial features like convolutional layers. Another advantage is the addition of new spatial memory, which can enhance the transportation of the spatial feature information in different layers. In order to test the model performance, two radars of Daxing District of Beijing and Guangzhou are analyzed. The radar echo is pre-processed through quality controlling to remove isolated echo, abnormal echo, invalid radial and echo below 15 dBZ and ground echo, and then the combined reflectivity (CR) is made by 0-5 layers of data. To examine the applicability of the PredRNN, a contrast experiment is designed between PredRNN and COTREC, including an independent verification over months of each radar and two severe convective cases analysis. The test is carried out by point by point in three different reflectivity thresholds:20 dBZ, 30 dBZ and 50 dBZ. Indexes of verification are CSI, POD and FAR. The time range of the test is 0-1 h by 6 min. Results show that PredRNN has better forecast performance in all the verification items especially in 20 dBZ and 30 dBZ, when the CSI can be raised by 0.15-0.30, POD can be raised by 0.15-0.25, and FAR can be reduced by 0.15-0.20. This effect of improvement enhances with time. Although forecast performances of both PredRNN and COTREC fall with time, the performance of PredRNN method descends more slowly. The forecast performances of both PredRNN and COTREC fall with the increase of the combined reflectivity factor strength, which shows the insufficient of prediction ability for the region with intensity over 50 dBZ. Two cases show that the PredRNN method has predictive ability for the change of reflectivity factor intensity. In summary, PredRNN is suitable for nowcasting, and its forecast performance is much better than COTREC.
-
Key words:
- nowcasting;
- recurrent neural network;
- deep learning
-
图 1 雷达资料预处理
(a)原始组合反射率因子, (b)预处理后的组合反射率因子
Fig. 1 Result of radar data pre-processing
(a)original composite reflectivity, (b)composite reflectivity after processing
图 3 2017年7月7日飑线过程实况和预报对比
(a)22：54北京大兴雷达组合反射率因子实况，(b)预测网络法21：54起报的60 min预报产品，(c)交叉相关法21：54起报的60 min预报产品
Fig. 3 Comparison between observation and forecast on 7 Jul 2017
(a)observation of composite reflectivity of Daxing radar in Beijing at 2254 BT, (b)60 min forecast at 2154 BT using PredRNN, (c)60 min forecast at 2154 BT using COTREC
图 4 2017年8月22日个例实况和预报对比
(a)21：30广州雷达组合反射率因子，(b)预测网络法20：30起报的60 min预报产品，(c)交叉相关法20：30起报的60 min预报产品
Fig. 4 Comparison between observation and forecast on 22 Aug 2017
(a)observation of composite reflectivity at 2130 BT, (b)60 min forecast at 2030 BT using PredRNN, (c)60 min forecast at 2030 BT using COTREC
表 1 数据集信息
Table 1 Information of dataset
站点 型号 学习数据集 独立检验数据集 北京大兴 SA 2014-04-01—10-31 2017-07-01—08-31 2015-04-01—10-31 2016-04-01—10-31 2017-04-01—06-31 广州 SA 2016-01-01—2017-07-31 2017-08-01—10-30 
下载: 导出CSV
表 2 北京大兴雷达检验集对比检验
Table 2 Quantitative result of Daxing radar in Beijing
检验指标 方法 30 min预报时效 60 min预报时效 20 dBZ 30 dBZ 50 dBZ 20 dBZ 30 dBZ 50 dBZ CSI 预测网络 0.63 0.43 0.14 0.52 0.32 0.05 交叉相关 0.41 0.27 0.04 0.30 0.17 0.01 POD 预测网络 0.78 0.59 0.23 0.70 0.47 0.10 交叉相关 0.64 0.45 0.09 0.53 0.33 0.03 FAR 预测网络 0.18 0.26 0.40 0.25 0.33 0.45 交叉相关 0.35 0.40 0.48 0.44 0.47 0.59
下载: 导出CSV
表 3 广州雷达检验集对比检验
Table 3 Quantitative result of Guangzhou radar
检验指标 方法 30 min预报时效 60 min预报时效 20 dBZ 30 dBZ 50 dBZ 20 dBZ 30 dBZ 50 dBZ CSI 预测网络 0.69 0.55 0.14 0.60 0.44 0.08 交叉相关 0.40 0.26 0.02 0.29 0.17 0.01 POD 预测网络 0.82 0.69 0.23 0.76 0.60 0.15 交叉相关 0.63 0.45 0.03 0.53 0.33 0.01 FAR 预测网络 0.15 0.20 0.37 0.20 0.26 0.40 交叉相关 0.30 0.37 0.49 0.41 0.44 0.50
下载: 导出CSV
表 4 2017年7月7日北京大兴雷达检验结果
Table 4 Quantitative result of Daxing radar in Beijing on 7 Jul 2017
检验指标 方法 60 min预报时效 20 dBZ 30 dBZ 50 dBZ CSI 预测网络 0.51 0.32 0.06 交叉相关 0.24 0.11 0.01 POD 预测网络 0.71 0.49 0.11 交叉相关 0.47 0.23 0.03 FAR 预测网络 0.29 0.35 0.45 交叉相关 0.48 0.51 0.75
下载: 导出CSV
表 5 2017年8月22日广州雷达检验结果
Table 5 Quantitative result of Guangzhou radar on 22 Aug 2017
检验指标 方法 60 min预报时效 20 dBZ 30 dBZ 50 dBZ CSI 预测网络 0.58 0.42 0.02 交叉相关 0.38 0.25 0.01 POD 预测网络 0.74 0.57 0.04 交叉相关 0.57 0.42 0.01 FAR 预测网络 0.21 0.27 0.43 交叉相关 0.34 0.39 0.50
下载: 导出CSV
-
[1] 陈明轩, 俞小鼎, 谭晓光, 等.对流天气临近预报技术的发展与研究进展.应用气象学报, 2004, 15(6):754-766. doi: 10.3969/j.issn.1001-7313.2004.06.015 [2] 石宝灵, 王红艳, 刘黎平.云南多普勒天气雷达网探测冰雹的覆盖能力.应用气象学报, 2018, 29(3):270-281. http://qikan.camscma.cn/jams/ch/reader/view_abstract.aspx?file_no=20180302&flag=1 [3] 陈雷, 戴建华, 陶岚.一种改进后的交叉相关法(COTREC)在降水临近预报中的应用.热带气象学报, 2009, 25(1):117-122. doi: 10.3969/j.issn.1004-4965.2009.01.015 [4] 张蕾, 魏鸣, 李南, 等.改进的光流法在回波外推预报中的应用.科学技术与工程, 2014, 32:133-137. http://d.old.wanfangdata.com.cn/Periodical/kxjsygc201432027 [5] Neill E H B, Clive E P, Alan S.Development of a precipitation nowcasting algorithm based upon optical flow techniques.J Hydrol, 2004, 288:74-91. doi: 10.1016/j.jhydrol.2003.11.011 [6] 郑永光, 周康辉, 盛杰, 等.强对流天气监测预报预警技术进展.应用气象学报, 2015, 26(6):641-657. http://qikan.camscma.cn/jams/ch/reader/view_abstract.aspx?file_no=20150601&flag=1 [7] Germann U, Zawadzki I.Scale-dependence of the predictability of precipitation from continental radar images.Part Ⅰ:Description of the methodology.Mon Wea Rev, 2002, 130:2859-2873. doi: 10.1175/1520-0493(2002)130<2859:SDOTPO>2.0.CO;2 [8] 王改利, 赵翠光, 刘黎平, 等.雷达回波外推预报的误差分析.高原气象, 2013, 32(3):874-883. http://d.old.wanfangdata.com.cn/Periodical/gyqx201303027 [9] 陈明轩, 王迎春, 俞小鼎.交叉相关外推算法的改进及其在对流临近预报中的应用.应用气象学报, 2007, 18(5):690-701. doi: 10.3969/j.issn.1001-7313.2007.05.014 [10] 孙志远, 鲁成详, 史忠植, 等.深度学习研究与进展.计算机科学, 2016, 43(2):1-8. http://d.old.wanfangdata.com.cn/Periodical/jsjkx201602001 [11] 余凯, 贾磊, 陈雨强, 等.深度学习的昨天、今天和明天.计算机研究与发展, 2013, 50(9):1799-1804. http://d.old.wanfangdata.com.cn/Periodical/jsjyjyfz201309002 [12] 陆虹, 翟盘茂, 覃卫坚, 等.低温雨雪过程的粒子群-神经网络预报模型.应用气象学报, 2015, 26(5):513-524. http://qikan.camscma.cn/jams/ch/reader/view_abstract.aspx?file_no=20150501&flag=1 [13] 卢新玉, 魏鸣, 王秀琴.TRMM月降水量产品在新疆地区的订正.应用气象学报, 2017, 28(3):379-384. http://qikan.camscma.cn/jams/ch/reader/view_abstract.aspx?file_no=20170311&flag=1 [14] 金莲文, 钟卓耀, 杨钊, 等.深度学习在手写汉字识别中的应用综述.自动化学报, 2016, 42(8):1124-1141. http://d.old.wanfangdata.com.cn/Periodical/zdhxb201608001 [15] Shi X, Chen Z, Wang H, et al.Convolutional LSTM Network: A Machine Learning Approach for Precipitation Nowcasting. NIPS, 2015: 802-810. [16] Li P W, Wong W K, Chan K Y, et al.SWIRLS-An Evolving Nowcasting System.Technical Note, 100.Hongkong Observatory, 2000. [17] Wang Y, Long M, Wang J, et al.PredRnn: Recurrent Neural Networks for Predictive Learning Using Spatiotemporal LSTMS.Advances in Neural Information Processing Systems, 2017: 879-888. https://papers.nips.cc/paper/6689-predrnn-recurrent-neural-networks-for-predictive-learning-using-spatiotemporal-lstms.pdf [18] Hochreiter S, Schmidhuber J.Long short-term memory.Neural Computation, 1997, 9(8):1735-1780. doi: 10.1162/neco.1997.9.8.1735 [19] Sutskever I, Vinyals O, Le Q V.Sequence to Sequence Learning with Neural Networks.NIPS, 2014: 3104-3112. https://papers.nips.cc/paper/5346-sequence-to-sequence-learning-with-neural-networks.pdf [20] Pascanu R, Mikolov T, Bengio Y.On the Difficulty of Training Recurrent Neural Networks.ICML, 2013: 310-318. https://arxiv.org/pdf/1211.5063.pdf [21] 韩丰, 魏鸣, 李南, 等.反射率因子和径向速度共同约束反演多普勒雷达风场.遥感学报, 2013, 17(3):584-589. http://d.old.wanfangdata.com.cn/Periodical/ygxb201303008 [22] 吴涛, 万玉发, 沃伟峰, 等.SWAN系统中雷达反射率因子质量控制算法及其应用.气象科技, 2013, 41(5):809-817. doi: 10.3969/j.issn.1671-6345.2013.05.004 [23] 孙赫敏.雷达回波中零度层亮带的研究.北京:中国气象科学研究院, 2014. [24] 曹杨, 陈洪滨, 苏德斌.C波段双线偏振天气雷达零度层亮带识别和订正.应用气象学报, 2018, 29(1):84-96. http://qikan.camscma.cn/jams/ch/reader/view_abstract.aspx?file_no=20180108&flag=1 [25] 韩丰, 沃伟峰.SWAN2.0系统的设计与实现.应用气象学报, 2018, 29(l):25-34. http://qikan.camscma.cn/jams/ch/reader/view_abstract.aspx?file_no=20180103&flag=1 [26] 谢超, 马学款.2017年8月大气环流和天气分析.气象, 2017, 43(11):1446-1452. http://d.old.wanfangdata.com.cn/Periodical/qx201711014 -
计量
- 摘要浏览量: 3791
- HTML全文浏览量: 1667
- PDF下载量: 416
- 被引次数: 0
