Forecasting Method Of Medium And Long-term Runoff In Upper Reaches Of Fenhe River

Runoff forecasting plays an important role in water resources management,utilization,protection and utilization.However,the formation of runoff is greatly influenced by human activities,climatic factors,watershed characteristics and other factors.As a result,runoff series has the complex characteristics of multiple scales,nonlinear,and non-stationary.Consequently,it is very difficult to accurately predict medium and long term runoff series.Therefore,in order to improve the accuracy of runoff forecasting,it is necessary to introduce better prediction methods,and to transform the runoff series into a stationary and simple one.The annual and monthly runoff data covering1958?2000 from four hydrological stations in the upper reaches of Fenhe River were adopted.Based on the analysis of runoff statistical and change characteristics,the applicability of single Echo State Network(ESN),Nonlinear auto-regressive Neural Network(NARNN),Random Forest(RF)and Long and Short Term Memory Network(LSTM)models,single decomposition hybrid models based on Complete Ensemble Empirical Mode Decomposition(CEEMD)or Extreme-point Symmetric Mode Decomposition(ESMD)and double decomposition hybrid models based on CEEMD-Variational Mode Decomposition(VMD)or ESMD-VMD is explored,so as to provide new methods and ideas for runoff forecasting.The main content and the main results of this research were as follows:(1)Intra-annual runoff distribution of four hydrological stations was very uneven,with the maximum volume of runoff in August,and mainly showing seasonal changes.After the 1970 s,the annual runoff distribution became more nonuniform.The inter-annual variation was large,and the overall trend was downward.Except for the Shangjingyou station on the tributary,the extreme value ratio of the other three stations increased from upstream to downstream,and the change was gradually drastic.Therefore,it can be preliminarily determined that the runoff series has non-stationary and nonlinear characteristics.(2)The Shangjingyou station showed a downward trend,but it was not significant,while the other three stations showed a significant downward trend.The annual and monthly runoff sequence of each station mainly had a mutation point around 1970 s,and the trend of downward mutation was consistent with the overall change.The annual and monthly runoff sequences contained more frequency information.The main frequency of the annual runoff sequence was0.07,0.14,0.21,0.28,0.32 and 0.42,while the main frequency of the monthly runoff sequence was 0.08,0.17 and 0.25.The amplitude of the annual and monthly runoff sequences was the maximum at the frequency of 0,these indicated that the runoff sequence contains noise information.In conclusion,it can be further determined that the runoff series has non-stationary,nonlinear and other complex characteristics.Augmented Dickey-Fuller(ADF)test was used to identify the non-stationary characteristics of the original runoff series.The results of ADF statistic values for the runoff series at four hydrological stations are greater than the critical value,which shows that the runoff series at these four hydrological stations are all non-stationary.Meanwhile,Fuzzy Entropy(FE)was calculated to determine the non-stationary and complex degree of the runoff sequence,in which the FE value of the monthly runoff series was between 0.93 and 1.26,and the FE value of the annual runoff was between 5.31 and 12.54.The results showed that the annual runoff series was more complex and non-stationary than the monthly runoff series.(3)When the length of validation period accounted for 80% of the total length,ESN,NARNN,RF and LSTM had the highest prediction accuracy.Therefore,the first 34 years and 413 months of runoff data were the simulation period,and the last 9 years and 103 months of runoff data were the validation period for modeling and prediction.Since each model has good nonlinear mapping ability,the Willmott's Index of Agreement(WIA)of the annual and monthly runoff forecasting results in the verification period was greater than 0.6,that is,the prediction was valid.Among them,the WIA value of RF and LSTM was relatively higher,namely,the combined regression model and deep learning model had better predictive performance.The accuracy of the annual runoff series was lower than that of the monthly runoff series,so it can be seen that non-stationary characteristic of runoff is still a complicating factor in forecasting.(4)The single decomposition hybrid models based on the CEEMD or ESMD were established.CEEMD and ESMD can effectively adaptivelydecompose a non-stationary time series into several stationary sub-series containing different frequency information.Therefore,both of the two single decomposition hybrid models can improve the prediction accuracy of runoff.The ESMD based hybrid models were better for the annual runoff series,and the WIA value was increased by 19%?23% on average compared with the single models during the validation period at four hydrological stations.As well as the CEEMD based hybrid models were better for the monthly runoff series,and the WIA value during the validation period was 19%?23% higher than that of the single models on average.However,the maximum frequency sub-series of the single decomposition hybrid models were still complex,Therefore,the prediction accuracy was still low.(5)The double decomposition hybrid models can decompose the maximum frequency sub-series into several simple Variational Mode(VM)by VMD,and further mine the information of the highest frequency sub-series,so as to improve the forecasting accuracy.The maximum frequency sub-series of CEEMD or ESMD was decomposed by VMD and then predicted,and the WIA value of the results could be increased by 6%?14% or 6%?10% in the annual runoff and 7%?16% or 9%?16% in the monthly runoff.So that the overall prediction accuracy can be effectively improved.For RF and LSTM methods with good prediction performance,combining them with the double decomposition to form the hybrid models,namely CEEMD-RF,CEEMD-LSTM,ESMD-RF and ESMD-LSTM,the prediction results are also better.Therefore,in order to improve the prediction accuracy of runoff,it is necessary to select effective runoff decomposition techniques and forecasting methods with better performance.