The Uncertainty Analysis Method Of Hydrological Simulation Based On MCMC And Its Application

It is of great significance for the optimal protection and management of water resources to effectively simulate and predict hydrological phenomena.However,the hydrological simulation process generally has lots of uncertainties,such as model input error,model structure error,parameters estimation error and so on.The MCMC method is a parameter estimation method based on Bayesian theory,and it's widely used in the parameters'uncertainty analysis of complex and high-dimensional hydrological models.Based on MCMC method,two methods of uncertainty analysis for hydrological simulation are proposed in this paper.?1?AR-MCMC method,a uncertainty analysis method which is combined with AR model and MCMC method,is used for parameter uncertainty analysis of deterministic hydrological model.It describes the autocorrelation of simulated residual sequence through AR model,so as to adjust the residual covariance matrix in traditional MCMC method.?2?The precipitation extremum simulation uncertainty analysis method which is combined with GEV/GPD and MCMC methods,is used to analyze the probability distribution characteristics and uncertainty of precipitation extremum.GEV and GPD distributions are used to construct the statistical model of precipitation extremum,and the probability density space of distribution parameters is identified by MCMC method.According to the above two methods,the snowmelt runoff of the Tizinafu river basin in Xinjiang and the precipitation extremums in Beijing,Shenzhen and Jinan were simulated.Through the comprehensive evaluation of the simulation performance of the two methods in practical application,the following conclusions are mainly obtained:?1?The MCMC method based on DREAM_ZS sampling algorithm can well identificate unknown parameters.For both of the deterministic hydrological model?SRM model?and the stochastic hydrological model?precipitation extremum statistical model?,it effectively identified the posterior probability distribution space of model parameters.The model obtained performs well,the Nash coefficient of SRM model in the recognition period and the verification period is 0.84 and 0.86 respectively,the correlation coefficient and deterministic coefficient of the statistical model of precipitation extremum are generally up to 0.95,the root-mean-square deviation is generally lower than 0.6,and the average absolute deviation is generally lower than 2.5.At the same time,the average width of the simulated prediction interval is moderate,the coverage of the data is high and the interval symmetry is good,which can effectively express the simulation prediction uncertainty caused by parameter uncertainty.?2?Compared with the traditional MCMC method,the AR-MCMC method is more helpful for taking better uncertainty analysis of simulation by the deterministic hydrological model,whose simulated residual sequences has obvious autocorrelation.Modifying the residual covariance matrix in the likelihood function based on AR model,can increase the convergence rate of uncertainty analysis,so as to improve the sampling efficiency of MCMC.In addition,compared with the traditional MCMC method,using the AR-MCMC method to take SRM simulation and uncertainty analysis can obtain a larger model edge likelihood value,the obtained model has better prediction performance,and the obtained runoff prediction interval has a higher inclusion rate and better interval symmetry to the observed data.?3?In the study of analyzing the characteristics of extreme precipitation events,a variety of extremum distribution functions and MCMC methods are combined to simulate different precipitation extremum sequences,is conducive to do a more comprehensive analysis of the probability distribution characteristics and uncertainty of precipitation extremum.The simulated confidence interval obtained by the GPD performs better,and the designed rainstorm of the same recurrence level obtained by the GEV is larger.As for different cities,the applicability of GEV and GPD is different,it's recommended to use a variety of distributions to simulate the extreme precipitation.