基于BP神经网络的大气污染物浓度预测

利用BP神经网络结合变量筛选的方法建立了SO_2、NO_2、O3、CO、PM_(10)、PM_(2.5)等6种污染物的浓度预测模型,并选取2014-01-01至2015-11-28时段,昆明市区5个环境监测点以上6种污染物浓度的监测数据建立了昆明市污染物日均浓度预测模型.采用平均影响值(Mean Impact Value,MIV)的方法筛选出分别对6种污染物日均浓度值有主要影响的变量,作为BP神经网络的输入变量,利用建立的预测模型分别对6种污染物的日均浓度进行预测.结果表明,在关上监测点利用浓度预测模型对SO_2、NO_2、O3、CO、PM_(10)、PM_(2.5)等6种污染物浓度进行预测,污染物浓度预测值和实测值趋势吻合度较高.变量筛选后SO_2、PM_(2.5)预测效果比变量筛选前的预测效果好.O3的均方根误差和PM_(10)的标准化平均偏差,变量筛选前的预测效果比变量筛选后的预测效果好.变量筛选前的NO_2和CO的预测结果比变量筛选后的预测效果好.其他4个环境监测点的污染物浓度预测结果与关上监测点的结果相似.

Forecast of air pollutant concentrations by BP neural network

In this study, six models for predicting SO₂, NO₂, O₃, CO, PM₁₀ and PM_2.5 concentrations were established by back propagation (BP) neural network and variable selection. They were applied in Kunming by making use of daily concentrations of the above pollutants measured from 2014-01-01 to 2015-11-28 at five stations. Furthermore, the mean impact value (MIV) method was used to identify major factors of daily concentrations of pollutants and the results were used as input variables of BP neural network to forecast daily concentration changes of pollutants. Analysis revealed that the prediction results of Guanshang coincided well with the observations. Moreover, the prediction results of SO₂ and PM_2.5 could be improved through variable selection, but the root mean square error of O₃, standard mean variation of PM₁₀, and NO₂ and CO predictions worsened. Similar performance was obtained for the other monitoring stations.