Forecast Of Dry Season Runoff From River Sources In Large Irrigation Areas In Central And Western Guanzhong

Accurate and real-time runoff forecasting of irrigation areas can provide scientific basis for irrigation water dispatch,which is of great significance for ensuring food security and promoting regional high-quality development.Aiming at the three irrigation area water source projects of Stone Reservoir,Wool Bay Reservoir and Fengjiashan Reservoir in western Guanzhong,the paper uses a variety of time series analysis methods to support each other,and comprehensively identify the evolution of river source runoff under changing environment.The runoff forecasting scheme of two time scales: total runoff in dry season and monthly runoff in dry season was studied.The main research contents and achievements are as follows:(1)Analysis of the interannual evolution characteristics of runoff Using Mann-Kendall test,linear regression test,moving average method,cumulative anomaly method,Bayesian integrated catastrophe test and power spectral density analysis and other methods to corroborate each other,the interannual evolution characteristics of runoff in the three reservoirs were comprehensively identified.The results show that: the years of the significant mutation of the river headwater runoff in the three reservoirs are different,but they all appeared in the 1980s;the whole sequence of the annual runoff shows a significant decreasing trend,but the decreasing trend of the yearly runoff after the unexpected change is not significant;The reduction range is between 27.3% and 40.9%,and the sudden change is the cause of the change in the yearly runoff trend;the three reservoirs have the same runoff cycle;the first cycle is about 12～14 years,and the second cycle is about 6～9 years.(2)Analysis of runoff variation characteristics within a year and division of flood and shrivelled seasons.According to the index of non-uniformity coefficient,concentration degree and concentration period to reflect the characteristics of the annual distribution of runoff,the Fisher optimal segmentation method and the monthly runoff reconstruction,hydrological variation analysis methods were used to divide the river source runoff of the three reservoirs into flood and shrivelled seasons.The results showed that the monthly runoff of the three reservoirs decreased to varying degrees before and after the mutation,ranging from 1% to56%.From the perspective of the uniformity of the annual distribution,the uneven coefficients of the runoff series of the three reservoirs all decreased after the mutation.The distribution within the year is more uniform than that before the sudden change,and the month with concentrated inflow is still August;the dry season in the study area is from November of the current year to March of the next year.(3)Simulation of total shrivelled season runoff.Using three decomposition methods,empirical mode decomposition(EMD),additional mode decomposition(VMD)and singular spectrum analysis(SSA),based on the long short-term memory network model(LSTM)optimized by the whale algorithm(WOA),the three reservoirs.The original shrivelled season runoff total sequence is simulated.The results show that suitable decomposition-simulation models can be selected for the three reservoirs for high-precision forecasting of the total runoff in the dry season of the next year.In the case of the same simulation model,the Stone River Reservoir recommends using the SSA decomposition model,with a regular pass rate of 92.8%,an accuracy level of Class A,a verification period pass rate of 85.7%,and an accuracy class of A;Wool Bay Reservoir is recommended to use EMD decomposition mode,regular pass rate is 91.7%,precision grade reaches Class A;verification period pass rate is 71.4%,meets the requirements of Class B program;Fengjiashan Reservoir recommends using VMD decomposition mode,regular rate and verification period pass rate are respectively 96.4% and100%,up to the requirements of the first-class program.(4)Monthly runoff forecast in shrivelled season.Using the combination of genetic analysis and statistical analysis to optimize the predictors of the monthly runoff in the shrivelled season,construct seven models based on multiple linear regression and machine learning methods,and forecast the monthly runoff in the shrivelled season of the three reservoirs respectively and evaluation.The results show that: using the precipitation in the same month,the precipitation in the previous month,and the runoff in the previous month as the predictors of the monthly runoff in the shrivelled season,the bidirectional long-short-term memory network model(Bi LSTM)can predict the monthly runoff in the shrivelled season in the Stone River Reservoir and Fengjiashan Reservoir The performance is the best among them,with regular pass rates of 86.4% and 74.7%,meeting the requirements of the first-class and second-class schemes respectively,and 71.4% and 62.9% of the verification period pass rates,respectively,meeting the requirements of the second-class and third-class schemes;The extreme learning machine model(KELM)performed best in the monthly runoff forecast in the shrivelled season of the Wool Bay Reservoir,with a pass rate of 80.8% in the regular period and 71.4% in the verification period,both of which met the requirements of the second-class plan.It is suggested that the two-way long-short-term memory network model be used for Shitouhe Reservoir and Fengjiashan Reservoir,and the nuclear extreme learning machine model be used for Maoaowan Reservoir to forecast the monthly runoff in dry season.