【作者】 郭俊；

【导师】 周建中；

【作者基本信息】 华中科技大学 ， 水利水电工程， 2013， 博士

【摘要】 流域水文预报是水文学研究领域的重要研究内容，对于水资源规划管理和可持续利用具有重要的支撑作用。在我国，独特的地理位置和气候条件决定了水资源时空分布的极度不均性，同时由于受人类活动和全球气候变暖的影响，当前的水资源时空分布特性及其演化规律正在发生深刻变化，特别是近年来随着我国大型水利枢纽工程的规划、建设和投运，流域水文时空变化特性日趋复杂，从而给流域水文建模与预报提出了更高的要求。本文围绕流域水资源开发及优化利用中面临的关键科学问题，研究流域水文建模与预报的先进理论与方法，研究工作对减低水旱灾害损失和实现水资源安全高效可持续利用具有重要的理论意义和工程实用价值。相关成果可在工程实践中推广应用，具有良好的工程应用前景。本文的主要研究内容和创新性成果包括：（1）针对集总式水文模型无法精细反映流域下垫面条件和流域水文循环各要素空间分布不均性，且难以描述水文系统特征相应的响应规律的问题，基于集总式新安江水文模型计算的基本原理，结合流域气候条件和下垫面时空变异规律，充分考虑流域下垫面空间分布的异质性和不同水文单元间的水平联系，将流域划分成若干个具有水平联系和垂向联系的栅格单元，并用严格的数学物理方程表述各栅格单元水文循环的子过程以及栅格单元间的水量交换关系，提出并建立了分布式栅格新安江水文预报模型。（2）基于两种典型的系统理论水文模型建模方法——神经网络和支持向量回归，建立了两种系统理论短期水文预报模型，研究结果表明，系统理论水文模型同样可以有效模拟流域水文系统径流过程的时间变化规律，支持向量回归模型表现出与神经网络模型相当的预报性能，有些典型条件下预报性能甚至优于神经网络模型，是神经网络模型的一种替代方法，同时，支持向量回归模型参数少，且相比神经网络模型建模步骤简单并具有较强地非线性拟合能力，具有良好的工程应用前景。（3）为求解水文模型率定的多目标复杂优化问题，提出了一种多目标文化混合复形差分进化算法MOCSCDE，MOCSCDE算法将SCE-UA置于文化进化的框架中，利用种群进化过程中提取的各种知识指导算法的搜索，提高算法的运行效率，同时考虑到SCE-UA中单纯形算子不能充分利用种群个体信息的不足，采用全局搜索能力强的差分进化算法替代单纯形算子，充分利用种群个体信息指导演化计算，进一步提高了算法的计算效率，为水文模型优化率定提供了一种高效的解决方法。（4）针对传统基于单一目标的水文模型参数优化率定方法不能充分全面挖掘水文系统不同动态行为特征的缺陷，将本文提出的多目标优化算法MOCSCDE用于集总式和分布式水文模型优化率定，有效避免了基于单目标优化率定产生的均化效应，显著提高模型预报性能，同时，将MOCSCDE算法应用于系统理论模型输入优选和模型参数优化，既可克服模型输入优选的难题，又能根据不同的水文特性自动优选不同模型结构，提高水文预报精度。（5）结合SCEM-UA算法分析了集总式和分布式水文模型参数不确定性和模型预报不确定性，同时首次证明了神经网络水文预报模型中存在的“过参数化”现象，并针对神经网络模型参数不确定性分析结果，提出了有效降低模型预报不确定性的方法，一定程度上提升了模型预报性能，提高了预报可靠性，并为全面评估模型预报性能提供决策依据。相关研究成果应用于国家科技支撑计划、国家973基础研究计划、水利部公益性行业专项等项目，为项目研究工作和工程实践运行提供重要的理论依据与数据支撑。更多还原

【Abstract】 Watershed hydrological forecasting is one of the most important issues in the field ofhydrology. It plays a vital role in water resources management and sustainable utilization.As a result of unique geographical position and the climatic conditions of China, spatialand temporal distribution of water resources is extremely uneven in our country. Also dueto the impact of human activity and global warming, temporal and spatial distributioncharacteristics and evolution of current water resource is undergoing profound change.Especially the construction of large water control projects is increasing the complex ofwatershed hydrological characteristics in recent years. This puts forward highrequirements on hydrological modeling and forecasting.In this thesis, the main attention is focused on watershed hydrological modeling andforecasting. This research can provide essential scientific support for optimal use of waterresources, and is of great theoretical significance and practical value for reducing floodlosses and achieving safe and efficient sustainable use of water resources. The main workand innovation of this thesis can be summarized as follows:(1) The lumped conceptual hydrological models cannot reflect the uneven distributionof watershed underlying surface and elements of the watershed water cycle. To overcomethis drawback, a novel distributed hydrological model named XAJGrid model is proposedin this research. XAJGrid model is an extension of classic lumped Xinanjiang model. Thisnovel model divides the watershed into several grids. The horizontal and verticalrelationship between each grid and its adjacent grids are described by physical equations.The XAJGrid model can fully take account of temporal and spatial variation of climateconditions and underlying surface.(2) Two typical system theoretic models are constructed based on Artificial NeuralNetwork and Support Vector Machine. The models’ feasibility is demonstrated throughseveral case studies. And the results show that both models can be capable of modelingthe dynamic characteristics of streamflow processes. Moreover, the Support VectorMachine model seems to be with better performance. Meanwhile, this model has strongnon-linear fitting capability and easy to implement. It can be seen as a valuable alternativeto Artificial Neural Network model.(3) A novel multi-objective algorithm named culture shuffled complex differentialevolution (MOCSCDE) algorithm is proposed to make model calibration based on multi-objective functions. In MOCSCDE algorithm, the culture evolution is taken as theevolving framework and the SCE-UA algorithm is employed as the core evolutionalgorithm for the population space. This evolution strategy can make use of the knowledgeobtained along with the evolution process to guide the algorithm toward the optimizationdirection. Meanwhile, the differential evolution algorithm is employed as a substitute ofthe simplex search operator in SCE-UA to enhance the efficiency of algorithm.(4) Traditional hydrological model calibration is always done with single objectivefunction. It cannot properly measure all characteristics of the hydrological system. Tocircumvent this problem, the proposed MOCSCDE algorithm is applied to optimize theparameters of hydrological models under the multi-objective optimization framework.This mechanism can significantly avoid the homogenization phenomenon existing intraditional hydrological model calibration. Besides, the MOCSCDE algorithm is adoptedto optimize the model structure and model parameters of a theoretic model based onSupport Vector Machine. This form of optimization can effectively improve the predictaccuracy by automatically choose the best model structure according differenthydrological characteristics.(5) The SCEM-UA algorithm is employed to analyze the parameter and predictiveuncertainty of lumped and distributed models. Simultaneously, the uncertainty of atheoretic model based on Artificial Neural Network is also discussed. The results showthat overparameterisation is very likely to be existed in this type of models. Furthermore,based on the results of uncertainty assessment, a method is proposed to reduce the modeluncertainty. It provides a new way of analyzing and reducing the uncertainty of theoreticmodels.Several relevant research results have been applied to National Science and TechnologySupport Program, National Basic Research Program of China (973Program) and SpecialResearch Foundation for the Public Welfare Industry of the Ministry of Science andTechnology. This can provide important theoretical basis and data support for theseprojects.更多还原