BP神经网络模型在水环境质量综合评价应用中的一些问题

Some aspects on application of BP neural network to comprehensive assessment of water environmental quality

BP神经网络是采用误差反向传播算法对网络权值进行训练的多层前向神经网络 ,以其优良的非线性逼近能力 ,获得广泛应用。建立的神经网络模型具有优异性能的必要条件是神经网络结构及其参数的合理选取、具有足够多和代表性、典型性好的训练样本、训练时求得全局最小点和不出现“过学习”或“过拟合”等。本文根据近几年BP神经网络建模理论研究的最新成果 ,研究发现目前在建立水环境质量综合评价BP神经网络模型时存在的几个主要问题 :①训练样本太少 ,②没有检验样本和测试样本 ,③神经网络结构太大等 ,从而极有可能造成在训练神经网络模型时出现“过拟合”或“过学习”现象 ,使建立的模型泛化能力较差或根本没有。在研究近年提出的应用BP神经网络方法建模的必备条件的基础上 ,对目前应用人工神经网络进行水环境质量综合评价的研究成果的分析表明 ,绝大多数水环境质量BP神经网络评价模型是在不满足建模条件的情况下建立的。计算实例表明 ,在不满足建模条件下建立的神经网络模型泛化能力和预测能力较差 ,极有可能出现多模式现象 。

BP neural network is a feed forward neural network that is learned according to error backpropagation algorithm. BP neural network with excellent nonlinear approximation ability is widely applied to various fields. The excellent nonlinear approximation ability of BP neural network is ensured by determining the topology and structural parameters properly, learning efficient training data set with good typical characters, searching the global minimum solutions and escaping overlearning during learning. According to the recently research results of BP neural networks modelling, some aspects on comprehensive assessment of water environmental quality using BP neural networks were presented and studied in this paper. The main problems are too few of training set data, no verification (validation) set data, too large in network topology, which thus resulted in overfitting and overlearning in training and poor generalization of the neural networks model set up. The necessary modelling conditions for BP neural networks were concluded. The many BP models, presented before, for water environmental quality were set up under the conditions inconsistent with the necessary modelling conditions. The case study shown that the model set up under conditions disagreement with the necessary modelling conditions possessed poor generalization, prediction capability, and possibly induced multimodal in connection weights.