Page 59 - 2023年第54卷第7期
P. 59

1869 - 1875.
                [34] VASWANIA,SHAZEERN,PARMARN,etal.Attentionisallyouneed[Z].arXiv,2017.
                [35] HUANGH,TANGQ,LIUZ.AdaptivecorrectionforecastingapproachforurbantrafficflowbasedonfuzzyC -
                       meanclusteringandadvancedneuralnetwork[J].JournalofAppliedMathematics,2013,2013(1):1 - 7.
                [36] LIT,QIANZ,HET.Short - Term loadforecastingwithimprovedCEEMDAN andGWO - Basedmultiplekernel
                       ELM[J].Complexity,2020.doi:10.1155?2020?1209547.



                         GroutingflowhybridpredictionmodelbasedonCEEMDAN - Transformer

                                   LIKai,RENBingyu,WANGJiajun,GUANTao,YUJia
                        (StateKeyLaboratoryofHydraulicEngineeringSimulationandSafety,TianjinUniversity,Tianjin300072,China)


                  Abstract:Groutingflowisoneofthemostimportantgroutingparametersofhydraulicengineering.Theabnormal
                  constructionconditioncanbefoundbytheeffectivegroutingflowpredictiontoguaranteetheconstructionqualityand
                  safety.However ,thegeologicalconditioniscomplexandgroutingflowdatahasthefeaturesofstrongnonlinearity
                  andvolatility,thereforethepredictionprecisionisunsatisfied.Theshortcomingsoftheexistinggroutingflowpre
                  dictionareasfollows :thetraditionalneuralnetworkmodelisinsufficientinfeatureextraction,resultinginunsatis
                  fiedpredictionprecision ;thetraditionalneuralnetworkmodelcalculatesoneresultbyonecalculation,multiple
                  timesteppredictionrequirescomplexmultiplecalculations;thepredictiontimeofonepointisshortandthepredic
                  tionresultcannotreflectthetotaltrendofgroutingflowsequence ,thereforeitisnotbeneficialtocontrolgrouting
                  flowandguaranteeconstructionquality.Forthoseproblems ,thisresearchproposesthegroutingflowhybridpredic
                  tionmodelbasedonCEEMDAN - Transformer.Thegroutingflowisdecomposedtoeigenmodefunctionandresidual
                  signalbasedonCompleteEnsembleEmpiricalModeDecompositionwithAdaptiveNoise(CEEMDAN),andthe
                  problemsofstrongnonlinearityandvolatilityaresettled.ThesequencepredictionofIntrinsicModeFunction (IMF)
                  isrealizedusingmulti - headattentionTransformer,andthetotaldependencybetweeninputdataandoutputdatais
                  establishedusingmulti - headattentionmethod.Thismethodiseffectiveinextractingdynamictemporalfeaturesand
                  improvingtheextractingquality.Finally ,thegroutingflowpredictionmodelwithmulti - inputandmulti - outputis
                  establishedtoimprovethecalculationefficiency ,providingthereferenceforgroutingflowcontrol.Theproposed
                  CEEMDAN - Transformermodelhasbettercalculationaccuracyandefficiencyingroutingflowprediction.
                  Keywords:groutingflow prediction; completeensembleempiricalmodedecompositionwithadaptivenoise;
                  Transformeralgorithm;attentionalgorithm;sequencetosequence

                                                                                    (责任编辑:李福田)





























                                                                                                —  8 1 7 —
   54   55   56   57   58   59   60   61   62   63   64