【作者】 焦树林；

【导师】 梁虹； 黄法苏；

【作者基本信息】 贵州师范大学 ， 自然地理学， 2003， 硕士

【摘要】 人工神经网络是一门正在迅速发展的交叉科学，现已广泛应用于机械工业、电子工业、能源工业、水利工程、化学工业、农业、环境和经济学等学科。本文主要讨论人工神经网络在喀斯特流域及其枯水径流研究中的应用。 影响喀斯特流域枯水径流的因素异常复杂，其影响因素主要有枯季前期降水量和岩性、地貌、流域面积、河网密度、主河道长度及其比降等流域下垫面状况。本文在前人研究的基础上，采用人工神经网络技术，结合传统的统计方法，研究贵州高原上喀斯特流域的枯水径流规律。 论文从影响喀斯特流域枯水径流的各影响因子入手，通过对入选流域的枯季前期降水量及其下垫面特征等16个因子的量化，从量上探求喀斯特流域的枯水径流各特征值与影响因子之间的相关规律。统计结果表明：流域的枯季前期降水、岩性、地貌、流域面积、河网密度、主河道长度及其比降等因子与喀斯特流域的枯水四特征的相关关系异常复杂。鉴于此，论文采用了主成分分析法，在基本保持原有信息的情况下，化高维为四维，提取了可包含有原始数据信息的85.345％的前四个主成分因子变量，并计算这四个主成分的因子得分。为了使样本具有代表性，论文利用自组织特征映射网络(SOM)可以在无教师示教的情况下，实现对输入模式的自组织分类的特点，以18个流域的主成分数据变量作为网络输入，在SOM(Self-Organizing Map)网络中进行了分类，把18个喀斯特流域分成三大类。第一类流域为黄猫村、下司、高车、下湾、官坝、高旺寨、皂角囤，它们的输出都为[1；0；0]；第二类流域为把本、对江、正安、荔波，它们的输出都为[0；1；0]；第三类流域为车边、石阡、旺草、土城、文峰塔、麦翁、平寨，它们的输出都为[0；0；1]。上述分类的结果综合反映了流域各个因子特征的最终结果，体现了流域影响因子对枯水的综合作用， 论文对比分析了分类前、后各因子与枯水特征的相关规律，结果表明，分类后各类流域中枯水特征与其各因子的线性相关程度明显提高。这充分说明了SOM网络对入选样本进行了有效的分类，使得聚类后各样本的共性增强，从而实现了样本具有代表性的目的。这摘要为建立基于四个主成分的分类枯水特征预测的BP(BaCk proPagat ion)模型提供了有力的依据。在基于流域因子的枯水特征预测模型的建立研究中，本文设计了5种类型的BP模型，用于预测枯水径流特征。这五种模型是:第一类是流域基于16个原始因子的BP预测模型，第二是分析流域基于主成分的BP预测模型，第三至五种模型分别为三类流域基于主成分的BP预测模型。在多次的试算过程中，本文对比分析了5类枯水特征预测的BP模型的特性，结果认为:采用主成分的BP模型可以大大提高网络的性能和预测能力，同时采用分类建立BP模型其枯水预测效果更佳。更多还原

【Abstract】 Artificial Neural Networks (ANNS) is a kind of cross-science, which is developing veryrapidly now. It has been used in many fields such as machinery industry, electronic industry, power industry, hydraulic engineering, chemical industry, agriculture, environment, economy and so on.This paper focuses on application of ANNS in the study of the karst basin and its low flow.There are many factors that work on the low flow complexly in karst basins. The factors include antecedent precipitation of low flow period and status of karst basins (such as lithology, landform, drainage area, drainage density, length and ratio of demotion of main river and so on.).Based on the former studies, integrated the conventional statistic methods, the technique of ANNS is applied in this paper for researching the law of the low flow of karst basins in Guizhou altiplano.In view of the factors of inflection of the low flow in karst basins, sixteen factors about antcedent precipitation and characteristics of the chosen karst basins, are quantified. And the correlation of low flow and the factors is worked out. The statistics show that the correlation of antecedent precipitation, lithology, landform, drainage density, drainage area, length and ratio of demotion of main river and four characteristics of low flow is very complex. In view of this, to maintain the information of the original data, the PGA (Principal Component Anylysis) method is applied to preprocess the input mode, which improves the generaziation performance of BP (Back Propagation) network by turning multi-dimension variable into four-dimension variable, and four principal components are take out containing 85.345% of the information of the original data. Based on the four principal components, the 18 chosen karst basins are classified into three types by the technique of SOM (Self-Organizing Map).The laws of correlation of unclassified karst basins and classified karst basins are compared. As a result, the grades of the correlation between factors and low flow are higher than those of the formers. In view of this, the forecasting models of BP are based on the four principal components.In the course of studying of the forecasting models of BP, the characteristics of five models are analysized carefully by contrast. It shows that the BP models based on four main factors may improve the performance and the capability of the network’s forecasting, meanwhile, it is very effective to set up forecasting model of BP based on four principal components of classified basins.更多还原