基于信号改进深度学习网络的京津冀城市群二氧化氮浓度预测

Prediction of Nitrogen Dioxide Concentration Based on Signal Improvement Deep Learning Neural Network in Beijing-Tianjin-Hebei Urban Agglomeration

在中国华北地区,二氧化氮污染仍旧不容忽视,尤其是在机动车辆密集和工业生产相对集中的京津冀城市群。运用小波分解(WD)和长短期记忆(LSTM)神经网络建立了W-LSTM组合模型,用于预测未来京津冀地区二氧化氮日均浓度和分指数。使用2014年1月—2018年5月主要大气污染物数据对组合预测模型进行训练试验,在获得最优模型参数后,使用2018年6月—2019年6月数据进行模型预测性能测试试验。结果表明,相较于传统的LSTM预测模型,W-LSTM组合预测模型具有更好的预测性能,预测结果的平均绝对百分误差为9.21%。在此基础上,使用最优预测模型对京津冀城市群2019年7月—2020年12月二氧化氮日均浓度进行了预测,并描绘了时空分布图用以表征其时空变化特征。

In North China,NO2 pollution still can’t be ignored,especially in Beijing-Tianjin-Hebei urban agglomeration with dense motor vehicles and relatively concentrated industrial production. In this paper,wavelet decomposition( WD) and long short-term memory(LSTM) neural network were used to establish a combined model named W-LSTM to predict the future NO2 concentration index in Beijing-Tianjin-Hebei urban agglomeration. The combined prediction model used the main air pollutant data from January 2014 to May 2018 for training experiment,and used the data from June 2018 to June 2019 for model prediction performance test experiment after obtaining the optimal model parameters. The results showed that the combined prediction model had good prediction performance compared with the single LSTM prediction model,and the mean absolute percentage error( MAPE) value of model evaluation index was 9. 21%. Finally,the optimal prediction model was used to predict the daily concentration of NO2 in Beijing-Tianjin-Hebei urban agglomeration from July 2019 to December 2020,and the spatiotemporal distribution map was drawn to reveal its spatiotemporal evolution characteristics.