CNN-LSTM PM2.5 Concentration Prediction Model Based On Attention Mechanism

PM2.5 adsorbed heavy metals,benzopyrene and other harmful substances will enter human lungs through the respiratory system,causing heavy metal poisoning,reproductive harm,and even cause cancer and other serious problems.Therefore,PM2.5 is widely regarded as an important indicator to measure the air quality of Chinese cities.Whether it is for people’s health,or the long-term development of the economy and the ecological environment,the advance prediction of PM2.5 concentration change is of positive significance.Based on the full analysis of domestic and foreign research methods for PM2.5concentration prediction,this paper builds the LSTM network prediction model and CNN-LSTM network prediction model,introduces the attention mechanism,designs and implements the CNN-LSTM prediction model based on self-training weight and fixed weight attention mechanism,and verifies the feasibility through simulation experiments.The main research contents of this paper include:(1)Aiming at the nonlinear time series phenomenon in the process of PM2.5 concentration change,an LSTM model suitable for processing time series data was built.Considering that the LSTM model only refers to the influence of time factors and ignores the spatial factors,which causes the model accuracy instability,the CNN network with strong capability in the field of feature extraction is introduced.Taking Hangzhou Xiasha Station and six neighboring stations as the experimental objects,by extracting the spatial correlation of neighboring stations except Xiasha Station to Xiasha Station and combining with the historical data of Xiasha Station as the input,the CNN-LSTM prediction model was built to obtain the PM2.5 concentration of the station in 1h,3h,6h and 10 h in the future.Through simulation experiments,it is concluded that the introduction of spatial information has a better prediction effect and interpretability for the singular value points that vary greatly in a short time.(2)In view of the decrease of model efficiency caused by the increase of data dimensions,a CNN-LSTM model based on the self-training attention mechanism is designed.In this model,the attention distribution vector α＿i was obtained by selecting the soft attention mechanism,defining the attention variable and setting the additive model as the attention scoring function,which was assigned to the hidden layer of CNN-LSTM model for training.Through the evaluation indexes of prediction accuracy,root mean square error(RMSE)and model efficiency,it is verified that the CNN-LSTM model performs better after the introduction of attention mechanism.(3)To introduce the CNN-LSTM model of attention mechanism,a method of fixing the probability vector of attention distribution is designed and implemented.In this method,KMO and Bartlett sphericity tests were carried out on the experimental data to determine whether it was suitable for principal component analysis,and the principal component analysis method in factor analysis was used to obtain the gravel diagram,total variance interpretation and component matrix,which were then put into the weight calculation formula of the comprehensive scoring model to obtain the weight of each adjacent site.To replace the probability vector of attention distribution in CNN-LSTM model based on self-training attention mechanism.Finally,the simulation results show that the model using this method performs better in RMSE and MAE,which proves the feasibility and effectiveness of this method in PM2.5 concentration prediction field.