基于BPNN的一江两河流域水体中重金属浓度预测

Prediction of heavy metal concentration in water of one river and two tributaries based on BPNN

本文将反向传播神经网络(BPNN)应用于青藏高原一江两河流域(雅鲁藏布江山南段、拉萨河、年楚河)水体中重金属浓度预测,探讨了输入变量、预测因子、隐藏层节点数和模型结构的影响.模型以溶解氧(DO)、pH、电导率(EC)、总磷(TP)、铁(Fe)作为网络的输入层,重金属砷(As)、锑(Sb)、钼(Mo)、锰(Mn)的含量作为网络的输出层,使用Levenberg-Marquardt(LM)算法进行训练.其中,BPNN隐藏层的传递函数为tansig,隐藏层节点数为9,输出层的传递函数为purelin,输出层节点数为4.结果表明:(1)以单个元素作为预测因子时,As、Sb、Mo、Mn预测值和实测值的决定系数(R2)分别为0.98、0.933、0.894、0.928;均方根误差(RMSE)分别为:9.7168×10^(−4)、1.2508×10^(−4)、3.3159×10^(−4)、1.9188×10^(−3).(2)以4个元素作为预测因子时,预测值和实测值的决定系数(R2)为0.888;均方根误差(RMSE)为2.1766×10^(−3).R2值越高,RMSE值越低,表明实测值和预测值拟合程度和适应性良好,证明BPNN能较好地应用于青藏高原一江两河流域水体中重金属浓度预测.

In this paper,backpropagation neural network(BPNN)is applied to the prediction of heavy metal concentration in water bodies of one river and two tributaries basins(the Shannan section of Yarlung Tsangpo River,Lhasa River and Nianchu River)of the Qinghai-Tibet Plateau,and the influence of input variables,predictors,number of hidden layer nodes and model structure are discussed.The model uses dissolved oxygen(DO),pH,electrical conductivity(EC),total phosphorus(TP)and iron(Fe)as the input layer of the network,and the content of heavy metals arsenic(As),antimony(Sb),molybdenum(Mo)and manganese(Mn)as the output layer of the network.Levenberg-Marquardt(LM)algorithm is used for training.The transfer function of the BPNN hidden layer is tansig and the number of nodes in the hidden layer is 9,and the transfer function of the output layer is purelin and the number of nodes in the output layer is 4.The results showed that:(1)When a single element was used as a predictor,the determination coefficients(R2)of the predicted and measured values of As,Sb,Mo and Mn were 0.98,0.933,0.894 and 0.928,respectively.The root mean square error(RMSE)was 9.7168×10^(−4),1.2508×10^(−4),3.3159×10^(−4)and 1.9188×10^(−3),respectively.(2)When the four elements were used as predictors,the determination coefficients(R2)between the predicted and measured values was 0.888.The root mean square error(RMSE)was 2.1766×10^(−3).The higher the R2 value,the lower the RMSE value,indicating that the measured and predicted values are well fitted and adaptable,which proves that BPNN can be better applied to the prediction of heavy metal concentrations in water bodies in one river and two tributaries basins on the Qinghai-Tibet Plateau.