BP人工神经网络在环境空气SO_2质量浓度预测中的应用

RESEARCH ON APPLICATION OF BP ARTIFICIAL NEURAL NETWORK IN THE PREDICTION OF THE CONCENTRATION OF SO_2 IN AMBIENT AIR

根据西安市雁塔区小寨环境空气监测点2011年7月31日起400 d的SO224小时平均浓度监测数据时间序列建立BP人工神经网络(ANN)预测模型,并用接下来100 d的数据对模型的仿真性能进行检验,从而验证了BP人工神经网络模型预测环境空气SO224小时平均浓度的可行性与准确度。经反复调试,最终选用2-3-1的网络结构并以trainbr作为训练算法,经34次迭代网络收敛,耗时7 s,预测结果相对于实际监测数据的平均绝对百分比误差为0.082,模型显示出良好的预测性能。预测结果表明,结构设定合理、训练算法选用适宜的BP人工神经网络模型能较好地反映SO2浓度的动态变化规律,具有可行性。

In order to verify the feasibility and accuracy of the back-propagation( BP) artificial neural network( ANN) model used in predicting the 24 h average concentration of SO2in ambient air,a BP artificial neural network( ANN) model was developed using the 24 h average concentration of SO2in 400 days,which was monitored on July 31,2011 at an environmental monitoring station in Xi' an. With the data in the next 100 days,the simulation performance of model was tested. After commissioning,'trainbr'was adopt as the training algorithm and the 2-3-1 network structure was used finally. After network convergence of 34 iterations in 7 s,the model showed a good prediction performance with MAPE = 0. 082( the mean absolute percent error between predictors and monitoring values). The prediction performance of the model indicates that a BP-ANN with proper network structure and training algorithm is able to reflect the dynamic variation of SO2concentration.