基于雷达回波的临近降水预报卷积循环神经网络模型研究  

Radar echo-based study on convolutional recurrent neural network model for precipitation nowcast

作　　者：王永灿 魏加华[1,2,3,4] 李琼 乔禛[1,2,3] 杨侃 朱旭东 包淑萍 王忠静 WANG Yongcan;WEI Jiahua;LI Qiong;QIAO Zhen;YANG Kan;ZHU Xudong;BAO Shuping;WANG Zhongjing(School of Water Resources and Electric Power,Qinghai University,Xining 810016,Qinghai,China;State Key Laboratory of Plateau Ecology and Agriculture,Qinghai University,Xining 810016,Qinghai,China;Laboratory of Ecological Protection and High-Quality Development of the Upper Yellow River Region,Qinghai University,Xining 810016,Qinghai,China;State Key Laboratory of Hydroscience and Engineering,Tsinghua University,Beijing 100084,China;Ningxia Meteorological Observatory,Yinchuan 750002,Ningxia,China;Ningxia Hydrology and Water Resources Monitoring and Early Warning Center,Yinchuan 750002,Ningxia,China)

机构地区：[1]青海大学水利电力学院,青海西宁810016 [2]青海大学三江源生态与高原农牧业国家重点实验室,青海西宁810016 [3]青海大学黄河上游生态保护与高质量发展实验室,青海西宁810016 [4]清华大学水沙科学与水利水电工程国家重点实验室,北京100084 [5]宁夏气象台,宁夏银川750002 [6]宁夏水文水资源监测预警中心,宁夏银川750002

出　　处：《水利水电技术（中英文）》2023年第1期24-41,共18页Water Resources and Hydropower Engineering

基　　金：国家自然科学基金项目(51909130);宁夏重点研发计划项目(2020BCF01002);水联网联合研究院项目(SKL-IOW-2020TC2004-02)。

摘　　要：【目的】临近降水预报是涉水灾害预警、洪水预报和调度管理等依赖气象预报决策的重要基础。高精度、高时空分辨率的气象雷达观测能够有效捕捉天气过程变化,发展基于雷达回波外推的临近降水预报方法,是中小流域高精度雨洪预报预警的关键。【方法】以银川贺兰山地区2017—2020年的C波段天气雷达和地面降水资料为基础,开展了ConvLSTM、ConvGRU和PredRNN三种卷积循环神经网络模型在不同降水情景下的预报性能研究,并将三种模型的预报结果与基于半拉格朗日外推的光流法进行对比分析。研究采用临界成功指数(CSI)、命中概率(POD)、虚警率(FAR)、均方根误差(RMSE)和结构相似性指数(SSIM)5种指标评估了三种模型在不同天气系统发展过程中的预报能力。【结果】结果显示:ConvLSTM模型可以较好的预测回波变化过程,而PredRNN模型对回波驻留和发展的过程预报效果较好;随着雨强的增大、预报时长的增加,卷积循环神经网络模型对回波运动的捕捉能力和回波强度变化的预测能力显著强于光流法;ConvLSTM模型能够更好的预报中小雨天气过程,结构更加复杂的PredRNN模型对暴雨过程具有更好的预报效果。【结论】结果表明:三种卷积循环神经网络模型中,ConvLSTM和PredRNN模型的预报效果优于结构较为简单的ConvGRU模型,且三种模型均优于光流法;在实际的应用中,1 h之内的预报可优先考虑ConvLSTM的预报结果,1 h后的预报则应更关注PredRNN模型的预报结果;三种卷积循环神经网络模型随预报时长的增加均出现“模糊化”“平滑化”的现象,需要从模型结构、训练方式等多方面进行改善。[Objective] Precipitation nowcasting is essential for weather-dependent decision making, such as early warning of water-related disasters, flood forecasting, and flood dispatching management. Weather radar observations with high precision and high spatiotemporal resolution can more accurately reflect the variations of weather processes. The development of a precipitation nowcast method based on radar echo extrapolation is crucial to high-precision rain, flood forecasts and early warning in small and medium-sized watersheds. [Methods] Based on the C-band weather radar and ground precipitation data in the Helan Mountain area of Yinchuan from 2017 to 2020, the forecast performance of three convolutional recurrent neural network models, ConvLSTM, ConvGRU, and PredRNN under different precipitation scenarios is studied. The prediction results of the three models are compared with the optical flow method based on semi-Lagrangian extrapolation. Critical success index(CSI), probability of hit(POD), false alarm rate(FAR), root mean square error(RMSE) and structural similarity index(SSIM) are used to evaluate the prediction ability of the three models in the development process of different weather systems. [Results] The results show that the ConvLSTM model can better predict the echo generation process, while the PredRNN model has a better prediction effect on the process of echo residence and development. The model’s ability to capture echo motion and predict the change of echo intensity is significantly stronger than that of the optical flow method. The ConvLSTM model can better predict the process of light and medium rain, and the more complex PredRNN model has a better predictive effect on the process of heavy rain. [Conclusion] Among the three convolutional recurrent neural network models, the forecasting effect of the ConvLSTM and PredRNN models is significantly better than the ConvGRU model which has a relatively simple structure. In practical application, the results of ConvLSTM can be given priority to within 1h

关 键 词：卷积循环神经网络 临近降水预报 雷达回波外推 深度学习 降雨 极端降水 气候变化 

分 类 号：P456.1[天文地球—大气科学及气象学]