Application Of Multitask Deep Learning In PM2.5 Prediction

Air quality prediction has been a general research interest and also a tricky research issue in the current field of environmental sustainability.In particular,prediction of PM2.5 can provide important importance for residents’ health and trip.In order to promote fine management of the environment,the government has deployed a series of micro-environmental quality monitoring stations to form intensive grid micro-monitoring station scenarios.At present,PM2.5 prediction studies mainly focus on sparse stations such as national base stations.Thus,research into intensive station scenarios is still insufficient.To study PM2.5 prediction in intensive station scenarios can promote the government’s fine management regionally.Compared with sparse stations,intensive stations has more compact spatial connection,of which the neighboring stations share temporal and spatial characteristics,thus providing new possibilities for applications of emerging artificial intelligence methods,such as deep learning and multitask learning,to PM2.5 prediction.This paper takes Lanzhou for a case study,and 359 intensive microscopic monitoring stations are adopted as the research objects to explore PM2.5 prediction methods within intensive station scenarios.Below is an overview of the research work in this paper:(1)Concerning insufficient research into PM2.5 prediction methods for intensive stations,this research puts forward the integrated PM2.5 prediction system for intensive scenarios based on deep learning,analyzes the optional application model,and provide methods for model training and assessment,thus laying a solid foundation for the followup research.(2)In response to the problem that the existing PM2.5 prediction methods cannot satisfy the application requirements of intensive stations,this paper proposes a mixed PM2.5 prediction model,CNN-GRU-FC,oriented towards intensive stations,and randomly chooses three stations from three regions in Lanzhou for analysis.The experiment compares the model with DNN,SVM,CNN,GRU,CNN-GRU in terms of PM2.5 prediction at the time node of “t+1”.Compared with other models,this model obtains the lowest MAE and RMSE in three stations.With CNN as the benchmark,the MAE and RMSE of the model in three stations improve by 31.94% and 24.93%,respectively.(3)The mixed model,CNN-GRU-FC,has a high degree of complexity,a large amount of training,and single-task learning,so it still has defects in digging spatial characteristics.A PM2.5 prediction model,MTD-CNN-GRU,based on multitask deep learning is put forward,which can learn other tasks as well.Under the same model construction conditions of CNN-GRU-FC,the experiment compares this model with CNN,GRU,and CNN-GRU in terms of PM2.5 prediction at the time node of “t+1” and “t+4”,respectively.Compared with other models,this model obtains the lowest MAE and RMSE in three stations at the time node of “t+1” and “t+4”,respectively.Similarly,CNN is adopted as the benchmark,and the model can improve MAE and RMSE of PM2.5 prediction at the time node of “t+1” by 39.51% and 28.52%,respectively.