BP神经网络地下水水质评价若干问题及修正

Problems of BP Neural Network in Groundwater Quality Evaluation and Amending

针对BP神经网络自身缺陷，利用Levenberg-Marquardt算法改善网络训练过程，并在地下水质量标准中采用随机插值生成样本，改进了网络训练样本过少和没有测试样本的问题。对三种样本获取模式下网络的训练结果进行对比，结果表明：在直接利用地下水质量标准作为样本模式下，网络无法收敛；利用linspace函数线性内插生成样本模式下，网络虽能很快收敛，但泛化能力差；利用rand函数随机插值生成样本模式下，当最优隐含层节点数为5时，网络正常收敛，测试样本相对误差均〈2％，网络泛化能力良好，精度较高。最后应用构建的最优BP神经网络模型对廊坊地区20个监测点的水质进行评价，评价结果与模糊综合评价结果基本一致，可信度高，能够反应该区地下水水质信息，可为相关政策制定提供重要依据。

For the shortcomings of BP neural network, Levenberg-Marquardt optimization algorithm was used to improve the training process. In order to improve the problems of small training samples and no test sample in the evaluation of groundwater quality, random interpolation method was used to generate sufficient training samples and test samples. Compared with three types of sample obtain model, the results show that： using groundwater quality standards as training samples, the network can not converge; using linear interpolation to generate training samples, although the network converge quickly, the network have bad generalization; using random interpolation to generate samples, When the number of hidden nodes is 5, the network converge normally, relative errors of the 20 test samples are less than 2%, and the net-work have good generalization and high precision. The optimal model was used to evaluate 20 groundwater monitoring samples of Langfang area and the results of which are consistent with the ones derived from fuzzy comprehensive evaluation. With high reliability, The results reflect the information of groundwater quality in these area and will be important foundation for constituting relevant environmental policies.