Some Studies On Hydrological Time Series Analysis

In this paper,hydrological time series is taken as the research object,and combined with entropy theory,copula,support vector machine and intelligent algorithm,change point analysis of hydrological time series;hydrological frequency parameter estimation and prediction model for runoff time series are studied respectively.In chapter 2,the change point problem of hydrological time series is discussed and studied.Firstly,the method of change point detection based on entropy theory is introduced into the change point analysis of hydrological time series.Secondly,random simulation method is used to analyze the performance of various entropy detection methods.In the end,the entropy detection methods are applied to the concrete instance.The results show that a series of entropy detection methods based on approximate entropy,sample entropy and permutation entropy are easy to operate.Without the help of other statistical tools,the position of the most likely change point can be accurately determined only by calculating the entropy value sequence.In addition,the entropy detection method has a small scale dependence on the sliding window and good robustness.The feasibility and practicability of the entropy detection method are further verified by applying it to the change point detection of daily runoff time series.Chapter 3 mainly studies the parameter estimation of P-? distribution.In order to improve the estimation precision,on the basis of in-depth study the advantages and disadvantages of GA,GA-RIM is proposed.Compared with GA,this algorithm has been improved in the three operations of selection,crossover and mutation.Especially it proposed the mutual exclusion mechanism of repeated individuals and the interpolation mechanism of variation,so as to keep the excellent individuals,increase the diversity of the population,accelerate the convergence speed of the algorithm and improve the precision of parameter optimization.Through the case analysis of the traditional parameter estimation method(moment method,the weight function method,probability weighted moment method,linear moment method,etc.),GA and GA-RIM,the results show that the parameter estimation method based on GA-RIM has higher precision and faster convergence speed in hydrological frequency parameter optimization of P-? distribution.In chapter 4,the annual runoff prediction problem of adjacent hydrological station is mainly studied.Firstly,three two-dimensional copula functions are deeply studied,and three prediction models based on single copula function are established,which are applied to the prediction of annual runoff time series.Secondly,on the basis of in-depth study of advantages and disadvantages of GA and PSO,PCGA-PSO is proposed.The algorithm is improved in the generation of initial population and the way of coding,the implementation and data fusion of the GA and PSO and the inertia weight setting of PSO.Through the case analysis of the traditional parameter estimation method,GA,PSO and PCGA-PSO,the results show that PCGA-PSO has the best performance in precision of parameter optimization.Finally,aiming at the existing problem of prediction models based on the single copula function,a hybrid copula function is proposed.PCGA-PSO is used to optimize parameters for the hybrid copula function.On this basis,an adjacent hydrological station prediction model based on the hybrid copula function is established and the model is applied to the prediction of annual runoff time series.The results show that the prediction precision and convergence speed of this model have obvious improvement compared with the prediction models based on the single copula function.Chapter 5 mainly studies the prediction of daily runoff with two variables.In order to predict the future daily runoff data according to the historical daily runoff data and daily precipitation data,an improved model based on least squares support vector machine is proposed.This model improves the traditional least squares support vector machine model in three aspects.(1)Instead of using a single kernel function,hybrid kernel function is adopted.The hybrid kernel function can take into account not only the generalization ability of support vector machine,but also its learning ability.(2)IPCGA-PSO is proposed on the basis of furthermore improvement of PCGA-PSO,and it is used to construct the hybrid kernel function.This algorithm not only improves the parallelism and data fusion,but also dynamically classifies the particles according to the fitness value of PSO,so as to dynamically change the inertia weight,avoid premature phenomenon and improve the convergence speed.(3)Penalty factor of the traditional least square support vector machine is redesigned so that it is no longer a constant,but a value that adaptively changes with the training process.Finally the rationality and reliability of the improved prediction model based on least squares support vector machine are illustrated by an example.In chapter 6,the insufficient of the study and some research prospects are provide on the basis of summarizing the whole paper.