Long Short Term Memory And Convolutional Neural Network（LSTM-CNN）for PM_2.5concentration Prediction

Accurate predictions of PM_2.5 concentrations can help people effectively avoid the harm caused by heavy pollution.Although China has gradually established air quality monitoring system,a large number of data from air quality monitoring stations and weather stations have not been fully utilized.Therefore,using historical data to predict PM_2.5concentrations is of great significance for government’s policy-making and people’s decision-making.Existing methods based on deep learning often emphasize the influences of local impact factors and neglect the effect of spatial transport,therefore,the predictions are usually unstable and the predictions of peak PM_2.5 concentrations are not accurate.In this paper,we propose a method,long-short-term memory-convolutional neural network（LSTM-CNN）,to predict PM_2.5 concentrations of a specific air quality monitoring station over 6 hours using historical data of PM_2.5 concentrations,historical data of weather conditions,and time stamp data.The correlations between PM_2.5 concentrations and local factors including meteorological factors and weather factors,and the correlation between PM_2.5 concentrations at different stations in a certain spatial range were analyzed.According to the correlation analyses,a LSTM-CNN composite model was constructed to predict the PM_2.5 concentrations of any monitoring station over 6 hours.The model is mainly composed of three parts:a)Using LSTM based time series prediction model to simulate local variations of PM_2.5 concentrations caused by local factors;b)Using CNN based predictive correction model to model variations of PM_2.5concentrations caused by spatial transport;c)Using FC based prediction aggregation model to dynamically combine the outputs of the above two models and obtain more accurate PM_2.5 prediction results.Finally,using ten-fold cross-validation to optimize the composite model.In this paper,Beijing-Tianjin-Hebei was chosen as the study area to verify the predicting performances of LSTM-CNN prediction model.A total of 21 central monitoring stations were selected from the Beijing-Tianjin-Hebei region as the research object,and the prediction performances of the proposed LSTM-CNN composite model were evaluated and compared with the traditional LSTM model.The prediction error and fitting degree of two models were compared and analyzed.The results showed that the LSTM-CNN model proposed in this paper had better prediction performance,better stability,and more accurate prediction of PM_2.5 peaks than the LSTM model.