基于前向神经网络的广义环境系统评价普适模型

为了建立由水环境、空气环境、生态环境、水资源环境、灾害环境、遥感环境、社会经济环境等不同环境系统组成的广义环境系统评价都能普适、通用的神经网络模型,针对BP神经网络因收敛速度慢、易于陷入局部极值而使实用性受限的缺陷,提出以双极性sigmoid函数作为网络隐层节点(神经元)的激活函数,而网络输出为所有隐层节点输出的线性求和的前向神经网络的广义环境系统评价模型.在设置广义环境系统指标参照值和指标值规范变换式,并对指标值进行规范变换的基础上,分别构建了适用于广义环境系统评价的任意2个指标规范值的前向神经网模型(NV-FNN(2)结构)和任意3个指标规范值的前向神经网模型(NV-FNN(3)结构).而对于指标较多的广义环境系统评价,只要将多指标分解为以上2个指标和3个指标的两种简单结构的前向神经网络的广义环境系统评价模型的组合表示即可.理论分析和实例检验结果表明:该模型对任意广义环境系统的规范指标值皆普适、通用,因而使不同环境系统的评价变得简洁、统一.规范变换和优化算法相结合的建模思想和方法对简化广义环境系统评价的多元回归、投影寻踪回归、回归支持向量机和径向基神经网络建模亦有借鉴和启迪作用.

Universal evaluation models for generalized environmental systems based on feed forward neural network

The purpose of this study is to build a universal and multipurpose neural network model for evaluating the generalized environmental system, which includes water environment, air environment, ecological environment, water resources environment, disaster environment, remote sensing environment, social-economic environment, etc. The traditional back propagation (BP) neural network model has limited usefulness because of its slow convergence and local optimum. Therefore, a feed forward neural networks (FNN) model was developed for evaluating the generalized environmental system. This model used the bi-sigmoid function as the activation function of hidden nodes, and its output was the linear sum of all the hidden nodes. This study set the proper reference values and the normalized transformation forms for all indexes of the generalized environmental system, and obtained the normalized values (NVs) of the indexes through normalized transformations. Furthermore, two types of NV-FNN models were built: the NV-FNN(2), which was suitable for the case involved any 2 normalized index values of the generalized environmental system, and the NV-FNN(3), which was suitable for the case involved 3 normalized index values. For the potential case with more than 3 indicators, it can be divided into several NV-FNN(2) and (or) NV-FNN(3) models.Our theoretical analysis and case studies showed that this NV-FNN model was suitable and universal to evaluate the normalized index values of any generalized environmental system, and it could make the evaluations of different environmental systems more concise and unified.The idea of the combination of normalized transformation and optimal algorithm, along with the corresponding methods, could provide reference for other model studies on the generalized environmental system, such as multiple regression, project pursuit regression, regression support vector machine, and radius basis function neural network.