基于变分自编码器的临近降水预报技术研究

打开文本图片集

Abstract：Idertodressteisuetatcuentnowastingpreipiatinmodelswereuable toanlinearfeaturevariatiosofrdar echoimagesinsatialndporalmesios，tsprpropodpreipiatioastigethodsdortioaltocder （VAE）a）Avariatioaltoencoderasiltodefieaprobabilitisributionfunciintelatentspaceforthebuldingofdaro images.b）Teslf-atentiomehanismasmploedtoleatedpendenisofadarchoimagesinspatialandtmpraldisios. c）Thediscretelatentspacewasiroducedtocapturethcomplexcontextualsemanticinformationofrdarechoimages.ThemodelpeforanceevaluationexperimentserecoductedonthSEVRdataset，withomparativenalysesagaisttetoepreseaiepriion nowcasting modelsofimpandPhyet.Thresultsindicatethathisodelachievesaccuratepredictionofradarchoimagesofenext5 frames，aditspreditioaccuracyisgherthanthatoftheimandPyDetmodels.Theintroductionofachmodulecancotributeto the performance improvement of the precipitation nowcasting model.

Key words：nowcasting of precipitation；deep learning；variational autoencoder；self-attention mechanism

0 引言

临近降水预报是一种短期天气预报技术，通常关注 0～6h 内的天气变化，其在气象学中具有重要应用[]。（剩余8695字）