人工智能模型的分类临近预报产品效果检验与分析

Effect Validation and Analysis of Classified Products Outputted by Artificial Intelligent Nowcasting Model

对2019年2—10月人工智能临近预报模型的降水、雷电、雷暴大风和冰雹等4类产品分别进行了业务检验与分析,结果表明:(1)对于降水产品,随着降水量级的增大,模型预报效果变差。相对来说,模型对0~1 h降水的预报能力略优于1~2 h,模型在0~2 h内完全不能预报出站点上≥10 mm的实况小时降水。(2)对于雷电产品,总体上模型的预报能力较差,1~2 h的预报效果明显比0~1 h差。相对5与20 km半径格点化实况,当取10 km半径格点化时,模型的雷电整体预报效果最优。(3)雷暴大风与冰雹产品具有类似的特点:在适当的时空匹配方式下,模型能预报出大多数实况中出现的雷暴大风或冰雹,但虚报率高,业务应用中主要考虑消空。最大时间提前量随着空间匹配半径或时间扩展的增大而增大,相对于冰雹预报,雷暴大风预报的最大时间提前量要优于冰雹预报。当空间匹配半径取10 km和时间扩展取20 min时,既可兼顾雷暴大风和冰雹的局地性,又可保证定量检验上相对较优。

Four products,namely precipitation,lightning,thunderstorm gale and hail,outputted by artificial intelligence nowcasting model from February to October in 2019,are verified and analyzed.It shows that,first,the performance of the Intelligent Nowcasting Model(INM)in precipitation prediction is getting worse with hour rainfall increasing.The hourly rainfall with truth column greater than or equal to 10 mm/h in 0-2 h could hardly be predicted by INM.The prediction accuracy within 0-1 h is slightly better than that for 1-2 h.Second,the model’s prediction effect for lightning are generally poor,the effect in 0-1 h significantly better than that in 1-2 h.Compared with the cloud-to-ground lightning detected by advanced time of arrival and direction system(ADTD)with 5 km and 20 km radius gridding,the overall lightning prediction effect of the model is relatively better for the 10 km radius gridding.Finally,INM performs similarly in predicting for thunderstorm gale and hail.Under the appropriate spatio-temporal matching,the model can predict mostly observed gales and hails accurately,and reduction the high false alarm ratio(FAR)for gale and hail should be mainly considered in operational application.The maximum leading time increases along with the rising of spatio-temporal matching radius or time.And the maximum leading time of thunderstorm gale is relatively much longer than that of hail.It is suitable for the local thunderstorm gale or hail and the relatively better performance in quantitative test with the space matching of 10 km radius expansion and 20 min time extension.