Study On Middle And Long Term Inflow Runoff Forecasting Of Reservoir Based On Hybrid Model And Copula Correction

As an important engineering measure of life and production,reservoir can alleviate the current situation of water resource shortage by optimizing operation.Reasonable planning and operation can not only prevent flood,but also meet the demand of interest.In order to maximize the comprehensive benefits of the reservoir and make reasonable planning and operation,it is necessary to make a scientific forecast of the upstream inflow,that is,to forecast the inflow runoff.The forecasting period of reservoir inflow runoff is more than three days and less than one year,so it is middle and long-term runoff forecasting.Accurate middle and long-term hydrological forecasting plays a key role in reasonable use of water resources.It is a powerful means and an important link to make full use of water resources and optimize reservoir operation.In this paper,the monthly runoff of Nierji reservoir from 1898 to 2012 is selected as the research object.The main research contents and conclusions are as follows:(1)In order to understand the basic characteristics and change rules of monthly runoff,in order to obtain more accurate forecast results,firstly,the statistical characteristics of monthly runoff in Nierji reservoir from 1898 to 2012 are analyzed,and its trend and stability are analyzed.The results are as follows: the monthly runoff of Nierji reservoir shows a weak growth trend in general;it shows an upward trend in flood season and a weak downward trend in dry season,but the downward trend is not obvious;the runoff distribution is asymmetric,and the extreme value is as high as 53.56:1;the monthly runoff changes are more intense in flood season and dry season,and the annual distribution is uneven,so it is difficult to predict.(2)The runoff time series are processed by the complementary ensemble empirical mode decomposition(CEEMD)and singular spectrum analysis(SSA),and the decomposition and reconstruction series are obtained.The least square support vector machine model(LSSVM)and the nearest neighbor sampling regression model(NNBR)are used to predict the obtained sequence.The hybrid model CEEMD-LSSVM,SSA-LSSVM and SSA-NNBR are constructed.For comparison,LSSVM and NNBR models are used to forecast the original sequence as input.It is concluded that the order of the performance of the model is CEEMD-LSSVM model,SSA-LSSVM model,SSA-NNBR model,LSSVM model and NNBR model.The hybrid model predicts that the runoff series and the observed series fit well,and are better than the single model,especially in the peak simulation.(3)Further correcting the forecasting results of the hybrid model.The Gamma distribution,Weibull distribution and Log-Normal distribution are used to fit the runoff forecast series,and the Normal distribution is used to fit the forecasting error values.By fitting the marginal distribution,the Frank-Copula function is used to establish the two-dimensional joint distribution of the monthly forecast values and error values of each hybrid model,and the conditional probability is calculated to obtain the conditional most likely value of the error under a certain forecasting value.Make corrections.The conclusions are as follows: the fitting effect of the marginal distribution is better,and the distribution of the predicted value distribution of different hybrid models has advantages and disadvantages;the joint distribution of the majority of the predicted value and the error value is better,and after the correction The NS values of the SSA-LSSVM model,SSA-NNBR model,and CEEMD-LSSVM model increased by 12.8%,18.3%,and 3.2%,respectively;the RMSE values decreased by 23.8%,16.1%,and 57%;and the MARE values decreased 38.9%,20.4% and 34.0%;the correction effect is positively correlated with the correlation coefficient between variables.(4)The forecast correction value from June to September in the flood season is used to calculate the relative error value.The four-dimensional Copula joint distribution is constructed by fitting the marginal distributions separately,and the possible values of the relative errors of multiple groups are randomly simulated,and then the combination of runoff and its corresponding occurrence probability in flood season are calculated,and the following conclusions are drawn: the fourdimensional t-Copula function can be better compared with the original relative error series,the overall distribution characteristics and distribution types of the random simulation series are better than those of the original simulation relative error series,and the gap is relatively small.This shows that the four-dimensional copula joint distribution of monthly runoff in flood season is credible.According to the copula joint distribution of monthly runoff in flood season,the probability of occurrence of the combination of monthly runoff forecast in flood season in the future years can be given,the probability forecast can be realized,and the basis of decision-making risk analysis can be provided for reservoir operation.