基于LSTM-GCN的PM2.5浓度预测模型

应用机器学习算法开展空气质量预测已成为当前研究热点之一,空气质量监测数据具有显著的时空特征,即具有时间维度时序特征和空间维度传输演化特征。面向空气质量监测数据,联合LSTM提取的时间特征和GCN提取的空间特征,提出预测PM_2.5浓度的LSTM-GCN组合模型。以北京市35个空气质量监测站2018—2020年监测数据进行仿真实验,并将LSTM-GCN模型与LSTM模型、GCN模型以及时空地理加权回归模型(GTWR)进行对比,结果显示:LSTM-GCN模型相较于LSTM模型均方根误差(RMSE)、平均绝对误差(MAE)分别降低了11.68%、7.34%;相较于GCN模型RMSE、MAE分别降低了40.22%、36.37%;相较于GTWR模型RMSE、MAE分别降低了17.52%、23.69%,表明所提出LSTM-GCN模型在准确率上有所提升。用LSTM-GCN模型预测2021年1—7月PM_2.5浓度,结果显示预测效果较好。

Prediction Model of PM2.5 Concentration Based on LSTM-GCN

The application of machine learning algorithm to air quality prediction has become one of the current research hotspots.The air quality monitoring data has significant spatial-temporal characteristics,that is,time-series characteristics and spatial evolution characteristics.Based on the air quality monitoring data,LSTM is used to extract temporal features and GCN is used to extract spatial features.A combined LSTM-GCN model is proposed to predict PM_2.5 concentration.The simulation experiment was carried out with the monitoring data of 35 air quality monitoring stations from 2018 to 2020 in Beijing,and the LSTM-GCN model was compared with LSTM model,GCN model,Geographically and Temporally Weighted Regression model(GTWR).The results showed that the model compared with LSTM model,Mean Square Root Error(RMSE),Mean Absolute Error(MAE) decreased by 11.68% and 7.34% respectively.Compared with GCN model,RMSE and MAE decreased by 40.22% and 36.37% respectively.Compared with GTWR model,RMSE and MAE decreased by 17.52% and 23.69% respectively.It showed that the LSTM-GCN model proposed in this study could effectively improve the prediction accuracy.Finally,the LSTM-GCN model was used to predict the PM_2.5 concentration from January to July 2021.The results showed that the prediction effect was satisfactory.