基于加权KNN-BP神经网络的PM_(2.5)浓度预测模型研究

PM_(2.5) prediction model based on weighted KNN-BP neural network

通过隶属度函数确定的加权KNN-BP神经网络方法,建立PM_(2.5)浓度动态实时预测模型,以PM_(2.5)、PM_(10)、NO_2、CO、O_3、SO_2等6种污染物前1 h的浓度及天气现象、温度、气压、湿度、风速、风向等6种气象条件,以及预测时刻所在一周中天数和该时刻所在一天当中的小时数为KNN实例的维度,选取3个近邻,根据得到的欧氏距离确定每个近邻变量的隶属度权重,最终将所有近邻的维度作为BP神经网络的输入层数据,输出要预测的下1 h PM_(2.5)浓度,该方法避免了传统BP神经网络方法不能体现历史时间窗内的数据对当前预测影响的问题。对北京市东城区监测站2014-05-01T00:00—2014-09-10T23:00的数据进行预测试验,结果表明,加权KNN-BP神经网络预测模型相较其他方法的预测误差最低,且稳定性效果最好,是PM_(2.5)浓度实时预测的有效方法。

Through the weighted KNN-BP neural network method determined by membership function, the dynamic real-time prediction model of PM2.5 concentration was established. The concentration of six pollutants,i. e. PM2.5, PM10, NO2, CO, O3 and SO2, six meteorological data including weather condition, temperature,pressure, humidity, wind speed and wind direction in the first hour, as well as the days of a week and the hours of the days for projection were regarded as the dimensions of the KNN instance. Three nearest neighbors were selected and, according to the Euclidean distance obtained, the membership weight of each neighbor point variable determined. Finally, the dimension of all nearest neighbor points were taken as the input layer of BP neural network, and the next hour PM2.5 concentration to be predicted as the output layer data. The method avoided the problem that the traditional BP neural network method failed to reflect the influence of the data in the historical window on the current predicting. The data of 2014-05-01 from 00: 00 to 23: 00 2014-09-10 in Dongcheng District monitoring station in Beijing was tested. The results showed that the prediction model with weighted KNN-BP neural network had the lowest deviation compared with other methods, and the stability showed the best. Therefore, this model is an effective method for the PM2.5 real time prediction.