贝叶斯规整化BPNN定量流域非点源污染负荷研究

将贝叶斯规整化BP神经网络(BRBPNN)应用于渭河流域非点源污染、社会和经济之间相互作用的研究。采用相关系数法确定输入变量为"降雨"、"种植地"、"草地"、"人口密度"和"羊密度",输出变量为总氮负荷。结果表明用BRBPNN定量非点源污染负荷是可行的,综合选择最优网络模型结构为BRBPNN(3c-7-1),其训练集和预测集相关系数分别为1.0000和0.9780,对应的均方误差分别为88.32和3.21×102。采用权值理论分析各输入因子对网络的贡献,依次为"降雨">"羊密度">"种植地"。该研究可为流域非点源污染的治理提供依据。

Modeling Changes of Non-point Source Pollution Load for Watershed Using Bayesian Regularized BP Neural Network

Bayesian regularized BP neural network (BRBPNN)was applied to quantify the impact of biophysical socioeconomy on non-point source (NPS) pollution within Weihe River. The input variables were rainfall,agricultural land, natural grass,population density and sheep density determined by correlation analysis(CA),and the output variable was TN load. Results showed that it was feasible to quantify NPS pollution load using BRBPNN. The obtained optimal network structure was 3c-7-1 and correlation coefficients were 1.0000 and 0.9780 for the training set and test set with 88.32 and 3.21× 102 of RMSE respectively. Network weight theory was applied to explain the relative importance of each input variable,with the decline order followed as rainfall,sheep density and agricultural land. The study can provide reference to NPS pollution management within a watershed scale.