Abstract:

In recent years, the artificial intelligence has made a breakthrough in image identification. In order to find out the practical value of artificial intelligence models in radar echo nowcasting in Wuhan City, the radar echo and precipitation observation data in Wuhan from 2015 to 2020 are used to train four deep learning models (PredRNN++, MIM, CrevNet and PhyDNet), then these trained models and radar echo observation data in flood season of 2021 are used to do nowcasting of radar echo. And on this basis, the precipitation processes are selected by using precipitation intensity and area indexes in Wuhan, and the performance of four deep learning models and optical flow method in radar echo nowcasting are tested and evaluated in Wuhan in flood season of 2021 in terms of mean square error (MSE), structural similarity index measurement (SSIM), probability of detection (POD), false alarm rate (FAR) and critical success index (CSI). The results are as follows: (1) On the whole, MSE of MIM model is the smallest, while its POD is the highest, and SSIM of MIM and PredRNN++ models are the highest. FAR of four deep learning models is lower than that of optical flow method, and it is the lowest for PhyDNet model. Except for CrevNet model, CSI of other three deep learning models is higher than that of optical flow method, and it is the highest for MIM model. (2) CSI of optical flow method is the highest during 0-12 minutes of forecast, while that of MIM model is the highest from 18 to 120 minutes, which shows the advantage of deep learning model for long prediction time. (3) With the increase of echo intensity, POD and CSI of four deep learning models and optical flow method decrease rapidly, while the variation characteristics of FAR of optical flow method and deep learning models are different. (4) For the regional precipitation processes, the prediction ability of deep learning models firstly reduces and then enhances significantly with the increase of precipitation intensity, while the optical flow method is insensitive to the change of precipitation intensity, so the increments of CSI of deep learning models are the highest under the strong precipitation processes compared with optical flow method. For the local convective precipitation processes with general intensity, the prediction ability of all models and optical flow method significantly reduces. (5) The analysis results of a rainstorm case show that deep learning models not only have prediction ability to the change of echo intensity to a certain extent, but also have better prediction ability to echo movement than optical flow method, so they have a good operational prospect.

Key words: deep learning models, optical flow method, radar echo, nowcasting, test and evaluation

摘要：

近年来，人工智能技术在图像识别领域取得了突破性进展，为探寻人工智能模型在武汉地区雷达回波临近预报中的应用价值，本文利用湖北武汉市2015—2020年雷达回波和降水量观测资料，对PredRNN++、MIM、CrevNet和PhyDNet 4种深度学习模型进行雷达回波临近预报训练，并基于2021年汛期雷达回波资料进行雷达回波临近预报。在此基础上，通过降水强度和降水面积指数筛选降水过程，并以均方误差（Mean Square Error， MSE）、结构相似性指数（Structural Similarity Index Measurement， SSIM）、命中率（Probability of Detection， POD）、空报率（False Alarm Rate， FAR）和临界成功指数（Critical Success Index， CSI）为指标，检验评估上述4种深度学习模型和光流法对2021年汛期武汉地区雷达回波的临近预报性能。结果表明：（1）整体来看，MIM模型的MSE最小、POD最高，MIM和PredRNN++模型的SSIM并列最高；所有深度学习模型的FAR均低于光流法，且PhyDNet模型的FAR最低；除CrevNet模型外，其余3种深度学习模型的CSI均高于光流法，且MIM模型的CSI最高。（2）预报的前12 min，光流法的CSI最高，而在18~120 min MIM模型的CSI最高，显示了深度学习模型长预报时效的优势。（3）随着回波强度增加，深度学习模型和光流法的POD和CSI均迅速降低，而FAR光流法与各模型则表现出不同的变化规律。（4）随着区域性降水强度增加，深度学习模型的预报能力均先降低后明显增强，而光流法对降水强度变化的敏感性较弱，故在强降水背景下深度学习模型的CSI较光流法增幅最大；对于局地一般对流性降水过程，所有深度学习模型和光流法的预报能力均大幅降低。（5）暴雨个例分析结果表明，深度学习模型不仅具备一定回波强度变化的预报能力，而且对回波运动的预报能力也明显高于光流法，展示了深度学习模型良好的应用前景。

关键词: 深度学习模型, 光流法, 雷达回波, 临近预报, 检验评估