Runoff Forecasting In The Uppper Reaches Of The Fenhe River Based On Combination Model

It is of great practical significance to predict runoff in the upper Fenhe River in water resources planning management,overall arrangement,Fenhe reservoir dispatching and flood control and disaster relief.The formation of runoff is influenced by human activities,climate change,rainfall,underlying surface and other factors.It is a nonlinear dynamic system with high complexity and weak dependence.The monthly runoff sequence is characterized by randomness and complexity,therefore the monthly runoff model with high precision is the difficulty and key point of runoff prediction.Monthly runoff data from 1958 to 2000 collected from Shangjingyou,Fenhe reservoir,Zhaishang and Lancun hydrological station in the upper Fenhe River were employed in this study.The monthly runoff data set from January1958 to December 1995 is used for training whilst that from January 1996 to December 2000 is used for prediction.The main contents and results of the study are as follows:(1)In order to explore characteristics of monthly runoff series at fourhydrologic stations in the upper Fenhe River,stationarity,trend and mutation of monthly runoff data were analyzed by KPSS(Kwiatkowski,Phillips,Schmidt,Shin)method,non-parametric Daniel test and Yamamoto method.The results show that the monthly runoff sequences at four hydrologic stations in the upper Fenhe River are non stationary ones,remarkably decreasing trends.In December1964,in January,February 1965,January 1972 and January,February,March1973,there was a mutation at Shangjingyou station;In November,December1971,January,February 1972,February,March 1973,there was a mutation at Fenhe reservoir station;In November,December 1971,January,February 1972,February,March 1973,there was a mutation at Zhaishang station;In November,December 1971,January,February,March 1972,January,February,March 1973,there was a mutation at Lancun station.In conclusion,the monthly runoff sequences in the upper Fenhe River present complexity and non-linearity.(2)Complementary Ensemble Empirical Mode Decomposition(CEEMD)can decompose complex time series,and then Intrinsic Mode Functions(IMFs)and one Residual(Res)with different characteristics can be obtained.At Shangjingyou hydrological station,the monthly runoff is decomposed by CEEMD approach into 8 IMF components and one residual.At Fenhe reservoir,Zhaishang and Lancun hydrological station,the monthly runoff is decomposed by CEEMD approach into 9 IMF components and one residual.CEEMD avoids the occurrence of mode aliasing compared with Empirical Mode Decomposition(EMD).(3)According to the nonlinear characteristics of monthly runoff series,Generalized Regression Neural Network(GRNN)has strong nonlinear fitting ability,CEEMD and GRNN combination model is used to forecast monthly runoff.Firstly,the monthly runoff is decomposed by CEEMD approach into multiple IMF components and one residual.Then GRNN are used to predict each component.Because the error comes mainly from the high frequency components,three data processing methods are used to reduce the error.The first method is to remove the high frequencies,the second is to add the linear coefficients for all components,the third method is to add the linear coefficients to the residual components after removing the high frequency items.The results show that the monthly runoff series at different hydrological stations are suitable for different modeling methods,and the prediction error of model 2 at Shangjingyou station is the smallest.The prediction error of model 1 at Fenhe reservoir station is the smallest.Model 3 at Zhaishang station and Lancun station has the smallest prediction error.The suitability at different hydrological stations for different models may be related to the variation coefficient of monthly runoff at hydrological stations.(4)Artificial neural network has advantage of solving nonlinear problems,CEEMD-GRNN-Elman neural network model,EMD-GRNN-Elman neural network model are used to forecast monthly runoff.Firstly,the monthly runoff is decomposed by CEEMD and EMD into multiple IMF components and one residual.In order to avoid errors from multi components,components should berestructured to high frequency component and low frequency component and trend.The GRNN was used to predict the high frequency and trend items,Elman neural network is used to predict the low frequency component.Results show that the prediction accuracy of CEEMD-GRNN-Elman neural network model and EMD-GRNN-Elman neural network model are higher than that of the single neural network model.CEEMD-GRNN-Elman neural network model is more accurate than EMD-GRNN-Elman neural network model,and can be used to predict medium and long term runoff.(5)Considering that mean generating function has the characteristic of periodic memory,CEEMD-mean generating function are used to forecast monthly runoff.First of all,the monthly runoff is decomposed by CEEMD.Then the high frequencies are removed to get the new sequence.Finally,mean generating function is used to predict new sequence.Results show that the prediction accuracy of CEEMD-mean generating function is higher than that of the single mean generating function.(6)Based on the above models,CEEMD-GRNN-Elman neural network has the highest accuracy and the best effect in predicting monthly runoff in the upper Fenhe River.