下击暴流智能识别和预报预警的方法研究

The downburst is a disastrous weather phenomenon, which is difficult to predict accurately in advance, and when it occurs, there is often a massive and fatal disaster. It is the most widely and feasible method to use radar for the identification and warning of the downburst. But some key features of the downburst in the radar are not clear，such as "radial positive and negative velocity" and "bow echo"，which make it extremely difficult to automatically identify and forecast the downburst..This paper mainly focuses on three aspects of the downburst: Intelligence identification, improving recognition accuracy and extending predictable scale. First, based on the morphological structure of the storm body in 3d space and its characteristics in the radar, the paper uses the graphics technology to build the identification model, and solves the identification problem of positive and negative speed..Second, using optical flow technology to build the time and space self-regression model based on optical flow vector field, solving the problem of inaccurate recognition of "non-rigid" target such as echo images by traditional optical flow method, improving the degree of precision of multi-dimensional prediction, such as location, time of occurrence and intensity. .Then, using the technology of deep learning to establish a feature recognition model for downburst in the early stage of formation. Through self-iterative training of historical meteorological data, mining intrinsic relationship of the physical quantity at the initial stage of formation of downburst. The trained model are able to make early recognition, prediction and immediate warning before the typical features have been identified by the radar and other detection data. .On the base of these, using heterogeneous parallel technology to build an algorithm model for the efficient and intelligent automatic identification tracking and prediction of the storm flow, reducing the computation of computer operation by an order of magnitude. Moreover, to gain more valuable time for the prediction and warning for downburst..The above research is able to promote the intellectual identification of early warning and forecast business critical flow level, enhance the veracity and reliability of the recognition, and extend the prediction time at 3 ~ 6 volume-scan times. The research findings can provide support that is more effective for forecasters, and provide more time for emergency measures such as evacuation of personnel and hazardous material transfer.

下击暴流是一种难以准确预测的灾害性天气。利用雷达等资料进行识别预警是当前较可行的方法，但其在雷达资料中所表现的一些关键性特征如“径向正负速度对”往往并不清晰，给自动识别和预报预警的业务化造成极大困难。.本项目围绕下击暴流的“智能识别”、“提高识别准确率”和“延长预报时效”三方面展开研究。首先以风暴体在空间的三维结构为基础，运用图形学技术构建识别模型，解决“正负速度对”等特征难以识别的问题。然后设计基于光流矢量场的时空自回归模型，改善传统光流法对回波图像这类“非刚体”目标识别不准确的问题。接着利用深度学习技术，构建下击暴流形成初期的特征识别模型，通过自我迭代训练深度挖掘其形成初期复杂的物理量内在关系，得到的模型能在雷达尚未出现“典型特征”前，提早做出识别和预警。最后运用异构并行技术改进下击暴流识别追踪和预报预警的算法，将运算时耗降低一个数量级。最终，为下击暴流的智能识别预警争取更多宝贵时间。

本项目的研究目标是对下击暴流做出及时、准确的识别和预报预警。研究过程从天气雷达图像模式识别、风暴体移动趋势追踪、下击暴流智能识别与预测三方面展开。.项目实施过程，首先运用了图像模式识别技术构建识别模型，解决“正负速度对”等回波特征难以自动识别的问题。然后采用光流法对反射率因子垂直剖面的光流场进行分析，得出风暴核心随时间演变的规律。接着采用拉格朗日力学模型对风暴核心顶高的下降过程进行函数拟合，再运用直方图和巴氏系数统计分析法，对风暴核心中层径向速度场中的“正负速度对”特征进行匹配识别，综合一系列阈值的判定，最终实现下击暴流的识别与预警。.实际研究内容与申请书中的技术路线、技术方法一致，并借鉴、融合了新近公开的一些新技术，如FlowNet、PredRNN++和MIM等基于AI技术的外推算法。算法改进主要有：提出基于多层迭代的局部约束光流算法，有效改善了传统光流法对回波这类非刚体移动目标的不适用性问题；运用深度学习等AI技术改善了雷达回波和径向速度外推预测的准确性，进而为下击暴流的识别、追踪和预警提供了更加准确的数据信息。利用GPU异构并行技术重构和优化算法，将模型的计算性能提升1个数量级。.研究过程分析检验了近12年的历史个例数据，结果表明，本套算法实现了下击暴流各项特征的自动识别，识别命中率（POD）达到78.9%~97.3%（不同个例的评分有差异），误报率（FAR）为22.7%~51.3%。误报率偏高的主要原因，一是对飑线等强对流天气的误判，二是深度学习模型的“数据不均衡”。如何在两项评分之间取得平衡，仍是本模型需要探究之处。总体而言，本研究实现了下击暴流0~30分钟预报准确率的显著提升，并将预报时效进一步延长2~3个体扫时间，可为预报业务人员提供辅助决策，为人员疏散、危险物转移等应急措施提供更加充足的时间。.研究形成的算法程序，现已部署在南京信息工程大学气象台测试运行。同时，已申请了国家气象信息中心的大数据云平台，预计2022年3月可在虚拟环境中读取实时数据试运行。