Study On The Prediction Of Hydrological Elements Time Series Based On The Data Extension And CEEMDAN Method

The evolution of rainfall,runoff and other hydrological elements is influenced by many factors,and under the joint action of various physical mechanisms in various subsystems contained in natural and social systems,the time series of various hydrological elements integrate microscopic and macroscopic fluctuation patterns and local and overall change characteristics,which contains rich available information,but also shows complex non-linear and non-smooth nature,which makes the evolution of hydrological elements The analysis of hydrological element evolution and hydrological prediction has brought difficulties.As a time-frequency analysis tool,CEEMDAN method can help us to analyze the rich information contained in complex hydrological variables and better grasp the internal evolution law of various hydrological elements.With its advantages of completeness and adaptability,the method has been rapidly promoted in the analysis and prediction research of time series of hydrological elements.However,CEEMDAN method itself will still be affected by the end effect,resulting in errors in the decomposition results.In this paper,the data continuation method is used to extend the original sequence to suppress the end effect,which is often ignored in previous studies.The decomposition results of the extended sequence were compared with those of the original sequence,and the improvement effect of the data continuation technology on the decomposition process was investigated.On this basis,the model construction idea of "extension-decomposition-prediction-reconstruction" was proposed,and the combined prediction models based on experience selection,parameter selection,weight optimization and mixed decomposition were built respectively to improve the prediction accuracy of time series of hydrological elements.The main research conclusions are as follows:(1)Data continuation and CEEMDAN decomposition were carried out for four groups of rainfall and runoff sequences in the source region of the Yellow River and Daqinghe Mountain,respectively,to explore the improvement effect of data continuation technology on the decomposition results.The results show that the decomposition accuracy of CEEMDAN method is affected by the end effect.Among the three commonly used data extension methods of mirror image extension,AR model extension and RBF neural network extension,the RBF neural network extension has the best effect and the end effect is effectively suppressed.The decomposition results can accurately reflect the fluctuation laws of different periods contained in the hydrological series.(2)In view of the problem that the "decomposition-prediction-reconstruction" model tends to ignore the end effect in the decomposition process,this paper proposes a combination prediction model construction idea of "continuation-decompositionprediction-reconstruction".Data continuation technology is used to suppress the end effect in the decomposition process to improve the decomposition accuracy and thus improve the prediction effect.Among the four combined prediction models constructed in this paper,the prediction accuracy of "extension-decomposition-predictionreconstruction" model is better than that of "decomposition-prediction-reconstruction" model,and the average relative error is reduced by 6.44%,2.94%,3.39% and 5.85%respectively.(3)In the process of "decomposition-forecast-reconstruction",a single model for component prediction cannot make full use of the different fluctuation information embodied by each component,which leads to the problem that the prediction results cannot accurately reflect the actual change law.Combined prediction schemes based on experience selection,parameter selection and weight optimization are proposed respectively.The prediction results of the components were optimized by selecting appropriate group prediction models with different methods to improve the prediction accuracy.The experimental results show that compared with the traditional single model prediction method,the three combined prediction models constructed in this paper can significantly improve the prediction effect,and the prediction accuracy is increased by 9.88%,6.13% and 8.90%,respectively.(4)This paper builds a combination prediction model based on "extension-decomposition-prediction-reconstruction",mixed decomposition and model selection.In the model,the strong nonlinear IMF1 component decomposed by the CEEMDAN method was further decomposed by wavelet to reduce the complexity of the data series and obtain a more stable component.The components decomposed by the two times were predicted and reconstructed respectively to obtain the final prediction results.Compared with the traditional single model and the combined model without wavelet decomposition,the average relative error of the prediction results is reduced by 10.51% and 5.85%.