多重组合神经网络模型在湖库总磷预测中的应用研究

APPLICATION OF MULTIPLE NEURAL NETWORK MODEL IN THE PREDICTION OF TOTAL PHOSPHORUS IN LAKE AND RESERVOIR

湖库总磷（TP）含量与环境因子的相关性往往并不显著，导致总磷预测精度不高，效果不理想。为提高总磷的预测精度，提出一种基于BP、Elman、RBF、GRNN（以下简称BP等4种）神经网络算法原理的组合预测模型，将影响总磷预测精度的NH4＋-N、CODMn和透明度3个相关因子作为BP等4种模型的输入向量，总磷实测值作为输出向量，构建3输入1输出的单一预测模型；以BP等4种单一模型预测结果作为下一BP等4种模型的输入向量，总磷实测值作为输出向量，从而构建4输入1输出的一次组合预测模型；再以一次组合神经网络模型预测结果作为下一BP等4种模型的输入向量，总磷实测值作为输出向量，构建4输入1输出的二次组合预测模型；依次类推，构建8种方案的多重组合预测模型。并构建GA—BP模型作为对比预测模型，预测结果与BP等4种单一模型及GA—BP模型的预测结果进行比较。结果表明：（1）组合模型随着组合重数的增加，预测精度显著提高，表明多重组合模型用于湖库总磷预测是合理可行的，模型具有较好的预测精度和泛化能力，是提高预测精度的有效方法；（2）方案2～8中各模型的预测结果均优于GABP模型（除方案2中的GRNN外），表明组合模型具有较高的预测精度和泛化能力。其中，方案3中的BP模型、方案4～8中的BP、Elman和RBF模型的平均相对误差均小于10％，预测精度均令人满意，尤以方案6～8中的BP、Elman和RBF模型的预测精度为最高（平均相对误差均在9％以内），均优于其他组合模型。

Total phosphorus （TP） concentration has often been found to do not correlate significantly to other environmental factors,leading to a low prediction accuracy of total phosphorus.In order to improve the prediction accuracy of total phosphorus, we proposed neural network algorithm combined forecasting models based on BP, Elman, RBF,GRNN （simplied as BP 4 in the following）.Setting the three related factors,i.e.NH＋-N,CODvn and transparency as the input and the measured values as the output,a single forecasting model of 3 input and 1 output was established.Then,using the output of BP 4 single model as the input of next BP 4 model,and total measured values as output, a combination forecasting model with 4 input and 1 output was established; taking this procedure once more and a secondary combination forecasting model with 4 input and 1 output was established, being followed by construction of 8 schemes of multiple combination forecasting model. The results showed as follows. In the combination models, the prediction accuracy is remarkably increased with increasing weight number of combination, indicating that multiple combination model for lake total phosphorus prediction is reasonable and feasible, and that the model has higher prediction accuracy and generalization ability, and it is the effective method to improve the prediction accuracy.All the predicted results from scheme 2 to scheme 8 were better than that from the GA-BP model （except for 2 GRNN）,indicating that the combination model has high forecasting accuracy and generalization ability.Among them, the average relative error of BP model in scheme 3, plan 4--8 in BP,Elman and RBF model was less than 10% ,demonstrating that the prediction accuracy is satisfactory.In the programme from 6 to 8,BP, Elman and RBF model, prediction accuracy is the highest （average relative errors are within 9 %）, being better than other combination model.