Research On Hydrological Model Development Methods Based On Optimal Sampling Of The Observation Data

In recent decades,the climate change,rapid urbanization,and water disasters occurred frequently,which seriously threatened the development of social economy,people’s personal and property safety.As an approximate description of the natural water cycle,hydrological models can effectively improve the people’s understanding of hydrological processes,and over the world.However,the accuracy of hydrological model prediction is affected by many aspects,including model structure,data noise,parameter optimizations and so on.Most researches focused on the structures and parameter evolution algorithms,while with little attention to the influence of data sampling,leading to the problems of poor generalization ability and insufficient prediction ability of hydrological models.Specifically,in the process of hydrological model construction,the observation datasets are usually divided into calibration and validation subsets,the former for model parameter optimization and the latter for evaluating the model performance.The existing data sampling methods are generally based on manual experience,which cause inevitable differences in distribution characteristics between the calibration data and the validation data.For example,the calibration dataset may contain a large number of hydrological extreme events,but the verified data set is all hydrologic normal values,resulting in a significant gap of performance between the model calibration and validation,causing poor model generalization ability,which significantly affects the engineering application of the hydrological models.To solve the above problems,two data sampling optimization algorithms has been developed to ensure that the calibration data and validation data have close distribution characteristics,thus improving the generalization ability of the hydrological models.The algorithms have been tested in the data-driven hydrological models.In order to expand the applicability of data sampling optimization algorithm in process-driven hydrological models,a new model calibration method has been proposed,which changes the traditional way that only continuous observation data should be used in process-driven hydrological models.Finally,based on the data sampling optimization algorithm,a new strategy for dynamic sampling and calibration of hydrological models has been established.The biggest feature of this method is that the distribution structure of the model parameters can be identified while the model parameters are calibrated,and the model uncertainty range can be determined.Firstly,based on the data-driven models,the basic characteristics of four commonly used data sampling optimization algorithms has been analyzed,and two new algorithms called SOMPLEX and MDUPLEX are proposed.These two algorithms have been used to construct data-driven models for 754 open catchments.The results show that they can significantly reduce the evaluation bias of the models,thus improving the generalization ability of data-driven hydrological models.Then,a hydrological model construction method based on discrete data is proposed,which drops the traditional model construction strategy of continuous hydrological observation data.The new method has been applied to three typical process-driven hydrological models and has been used in 163 open catchments.The results show that the discrete data model construction method proposed in this study can significantly reduce the gap between calibration and validation performance,and thus improve the generalization ability of process-driven hydrological models.Finally,combined with the data sampling optimization algorithms,a dynamic sampling strategy has been adopted during the process-driven hydrological model calibration,which means that the model calibration algorithm constantly updates the calibration data during the evolution process,thus identifying the distribution characteristics of the parameters while the model parameters are calibrated.Compared with the traditional fixed-data calibrating strategy,the model parameters can converge to a regional distribution,which is helpful to identify the uncertainty range of the model prediction.The new strategy has been tested in 163 catchments,and the results show that the proposed strategy can improve the performance of the hydrological models,and provide the range of prediction uncertainty.