Page 76 - 2024年第55卷第5期
P. 76

toimprovethepredictionofweeklyCOVID - 19casesinGermany[J].ScientificReports,2022,12(1):3930.
                [24] MEIJ,MA G, WANG Q, etal.Micro - and macroscopicaspectsoftheintermittentbehaviorsofgranular
                       materialsrelatedbygraphneuralnetwork[J].InternationalJournalofSolidsandStructures,2022,251:111763.
                [25] MAZ,MEIG,PREZIOSOE,etal.Adeeplearningapproachusinggraphconvolutionalnetworksforslopede
                       formationpredictionbasedontime - seriesdisplacementdata [J].NeuralComputingandApplications,2021,33
                       (21):14441 - 14457.
                [26] 张振东,罗斌,覃晖,等.风 光 水 互 补 系 统 时 间 序 列 变 量 概 率 预 报 框 架 [J].水 利 学 报,2022,53(8):
                      949 - 963.
                [27] SALINASD,FLUNKERTV,GASTHAUSJ,etal.DeepAR:Probabilisticforecastingwithautoregressiverecur
                       rentnetworks [J].InternationalJournalofForecasting,2020,36(3):1181 - 1191.
                [28] 张艺,迟世春,周雄雄.基于统计模型的高心墙堆石坝沉降变形研究 [J].水利与建 筑 工 程 学 报,2019,
                      17(4):141 - 146.
                [29] 程光,钱德鑫,郭建伟,等.基于散 度 的 网 络 流 概 念 漂 移 分 类 方 法 [J].计 算 机 研 究 与 发 展,2020,57
                       (12):2673 - 2682.
                [30] LIM,SHENY,REN Q, etal.A new distributedtimeseriesevolutionpredictionmodelfordam deformation
                       basedonconstituentelements [J].AdvancedEngineeringInformatics,2019,39:41 - 52.



                       Spatial - temporalfusionmodelfordeformationpredictionofrockfilldamsand
                                            itsapplicationinsafetymonitoring

                                    1
                                               1,2,3
                                                                                         1,2,3
                                                                          1,2,4
                                                             2
                             WUJiye,MAGang        ,AIZhitao,YANGQigui ,ZHOUWei
                      (1.StateKeyLaboratoryofWaterResourcesEngineeringandManagement,WuhanUniversity,Wuhan 430072,China;
                                2.InstituteofWaterEngineeringSciences,WuhanUniversity,Wuhan 430072,China;
              3.KeyLaboratoryofRockMechanicsinHydraulicStructuralEngineeringofMinistryofEducation,WuhanUniversity,Wuhan 430072,China;
                                            4.CISPDRCorporation,Wuhan 430010,China)
                  Abstract:Deformationpredictionisthekeyforsafetymonitoringandhealthassessmentforrockfilldams.Current
                  researchmostlyfocusesonsingle - pointdeformationpredictionmodels,neglectingthemulti - pointcorrelationfor
                  theoverallmodeling.Besides ,itischallengingforcurrentmodelstoachievelong - termaccuratepredictionofdrift
                  deformationdata.Consideringtemporaldependenceoftimeseriesandspatialcorrelationbetweenmultipointforthe
                  deformationofrockfilldams ,aspatial - temporalfusionmodelbasedonGraphConvolutionalNetwork(GCN)and
                  RecurrentNeuralNetwork (RNN)isproposedfordeformationprediction,introducingprobabilisticpredictionand
                  full - processtraining.Firstly ,themodeladaptivelyperformsmultipointfeaturesfusionusingGCN.Then,the
                  transmissibilityofcellstatesandhiddenmemoriesalongthetimeaxisinRNNisutilizedtorealizetheminingand
                  fusionofspatial - temporalinformation.Finally ,theparametersoftheprobabilisticpredictionareobtainedaslinear
                  layeroutputtoimprovethemodel ’srobustnessagainstnoiseinmonitoringdata.Inordertoenhanceitsabilityto
                  understandtheintrinsicrelationshipbetweeninfluencingfactorsandcumulativedeformation,themodeladoptsa
                  full - processtrainingandinferencetechnique,whichrealizeslong - term accuratepredictionfordriftdeformation
                  data.TakingShuibuyaconcrete - facedrockfilldam asastudycase ,weconductcomparisonandablationexperi
                  ment,thenpresentthreespecificapplicationsofthismodelinsafetymonitoringandhealthassessmentforrockfill
                  dams.Ourresultsdemonstratethatthemodelsuccessfullyintegratesthespatial - temporalinformation ,significantly
                  improvingpredictionaccuracycomparedtocurrentmodels.Itaddressesthechallengesoflearningthegenerallaw
                  properlyandpredictingdriftdeformationdataaccuratelyofrockfilldams ,andcanbeappliedforlong - termdeform
                  ationprediction ,anomalydetectionandmissingdatacompletionofmeasurementpoints.
                  Keywords:rockfilldams;deformationprediction;spatial - temporalfusion;GraphConvolutionalNetwork;Long
                  andShort - TermMemoryNetwork ;probabilisticprediction


                                                                                    (责任编辑:韩 昆)


                —  5 7  —
                     6
   71   72   73   74   75   76   77   78   79   80   81