【作者】 郭劲松；

【导师】 龙腾锐；

【作者基本信息】 重庆大学 ， 市政工程， 2002， 博士

【摘要】 人工神经网络(ANN)是复杂非线性科学和人工智能科学的前沿，其在水污染控制规划领域的应用研究在国内外尚处于初创阶段。本文在较全面分析评述了水质评价与水质模拟研究现状，及在分析阐述了ANN基本原理、算法和各类模式特征的基础上，在国内首次将ANN方法引入水污染控制规划领域，主要在水质评价和水质模拟的人工神经网络建模方法以及模型算法方面进行了一些创新性的工作，为提高水质评价和水质模拟的智能化水平做出了努力。 本文根据水质综合评价的特点及Hopfield网络优良的模式识别性能，通过对Liapunov能量函数构造的合理设计，提出了水质综合评价Hopfield网络模型，并从数学上严格推证了水质评价Hopfield网络的样本分类性能。实例研究表明：Hopfield模型在相当多评价指标的情况下，仍可很快地给出评价结果，且可达到相当高的精度，同时模型可表述定量和定性的评价指标，增强了评价方法的通用性和适用性。 针对水质信息的模糊性特征，本文将模糊数学与神经网络相结合，首次提出了水质评价隶属度BP模型。通过对隶属度BP模型、模糊综合指数法和灰色聚类法实例评价结果的比较，本文提出的隶属度BP模型融合了ANN方法和模糊评价方法的优点，有效地克服了模糊综合指数法评价结果偏重和灰色聚类法评价结果偏轻的缺陷，提高了评价结果的准确性和可靠性；该模型可方便地对模糊规则进行增加或删减，比传统的综合指数类评价方法更具灵活性，程序的通用性好，应用方便；隶属度BP模型考虑了环境水质类别变化的连续性，使评价方法更接近客观实际。 本文根据对污染物在河流中迁移过程分段传递特征的分析，首次提出了模拟污染物在河流中迁移规律的串联ANN模型，该模型是对ANN模型结构的一种改造。根据这一建模思路，建立了BOD-DO耦合BP网络水质模拟模型，以及学习速率采用DBD自适应技术的一维水质综合模拟BP网络模型。实例研究表明：ANN模型对水质模拟的结果比一维水质模型模拟的结果精度更高，验证了本文提出的串联ANN模型结构的有效性和正确性。本论文在第8章提出了优化构造算法的一维水质综合模拟RBF网络模型，为克服BP网络在最小值附近容易产生振荡的不足，以及利用训练样本解决网络结构的自适应优化等问题进行了有益的探索。实例研究表明：该RBF模型具有更高的水质综合模拟泛化能力，提高了ANN水质综合模拟的预测精度，展示了ANN水质模拟模型具有良好的应用前景。 根据二维水质数学模型数值解法的思路，本文开创性地提出了进行二维水质模拟的广义网络法。该方法根据二维水质扩散三个不同方面问题的特征，采用前馈网络的拓朴结构形式，以河段特征参数神经元直接表达、横向扩散系数三层BP网络 重庆大学博士学位论文表达、纵向扩散系数经验公式表达等构造了不同表达方式的输入层神经元，进而依不同表达方式的输入神经元构造了广义网络。通过实例研究验证了该ANN二维水质模拟方法的准确性与适用性。 本文研究表明：用ANN模型来进行水质评价和水质模拟或预测在理论上可行，在实践上有继续深入研究开发的价值，具有良好的应用前景。本研究为水质模拟研究提供了一种新方法，开辟了一条较好的新途径，也为ANN的应用增添了新领域。更多还原

【Abstract】 Innovative work is done on modeling and algorithm of artificial neural networks (ANN) for water quality assessment and simulation in this dissertation.Artificial neural network plays a leading role in the sciences for complex non-linear phenomena and artificial intelligence. Researches on its application in the planning of water pollution control are still in the preliminary stage in the world. On the basis of a comprehensive evaluation and analysis of the present situation of the researches in water quality assessment and simulation, and on the basis of a careful exposition of the basic principles, the algorithm and the varied pattern features of ANN, this dissertation gives an application of ANN approaches in water quality assessment and simulation, which, as the first attempt of its kind, can help to achieve a higher level in the application of artificial intelligence in this field.Based on the features of integrated water quality assessment, and the outstanding pattern recognition capacity of Hopfield network, a rational design of the structure of Liapunov energy function, this dissertation proposes the first Hopfield network model for comprehensive water quality assessment, and gives a strict mathematical deduction of the sample-classification performances of this model. Case study reveals that with numerous assessment indexes, Hopfield model can quickly produce assessment results with a high accuracy, and, as it can work with both qualitative and quantitative indexes, has wider areas of application.This dissertation combines Fuzzy Mathematics and ANN and produces a membership degree Back-Propagation network (MDBP) for water quality assessment compatible with the fuzzy features of water quality data. Case studies of the results of this model approach, fuzzy integrated index approach, and grey accumulation approach indicate: the proposed MDBP model combines the merits of ANN approach and fuzzy approach, and improves the accuracy and reliability of the assessment results while the results of fuzzy colligation index approach and grey accumulation approach are too high and too low respectively; it has a higher flexibility than conventional integrative index approach and its programs have a better adaptability and more convenient application; its ways of assessment are closer to the reality since it takes into consideration the continuity of the changes of environment water quality.Based on an analysis of the section-transmission features of the transference of thepollutants in the rivers, this dissertation creates a series ANN model, an improved ANN model, to simulate the transference of the pollutants in the rivers. In addition, it establishes a model of BOD-DO coupling for water quality simulation based on BP network, and a one-dimensional BP network model with its learning rate using Delta-Bar-Delta (DBD) self-adaptation technology. Case studies show a higher accuracy in the simulated results of ANN model over the one-dimensional model, thus proves the effectiveness, correctness and a bright future for the application of the proposed series ANN model structure.To overcome the tendency towards oscillation of the BP network near the minimum value, Chapter 8 proposes a one-dimensional radial basis function (RBF) network model for comprehensive water quality simulation for optimizing structure with algorithm, which is a useful exploration into issues such as the self-adaptive optimization of network structure through training samples. Case studies show that this RBF network has a higher generalization capacity in water quality simulation, thus improves the prediction precision of comprehensive water quality simulation with ANN.The numerical value solution of the two-dimensional mathematical model for water quality inspires a broad-sense network method for two-dimensional water quality simulation. This method takes into consideration the characteristics of the three aspects in two-dimensional water quality dispersion, and constructs input neurons of different expressions, the topological structu更多还原