Research On PM2.5 Concentration Prediction Based On Attention Mechanism And CNN-GRU

In recent years,the problem of air pollution has also become increasingly serious and a series of problems have arisen as a result,causing many problems to people’s productive lives and health.PM2.5,one of the most important pollutants in air quality,can cause a certain degree of harm to the human body.Therefore,a reasonable prediction of PM2.5concentration and good preventive and protective measures can effectively safeguard the health and safety of the body,as well as the long-term ecological development.In this paper,based on a thorough understanding and analysis of domestic and international research methods for PM2.5 prediction,a hybrid prediction model of gated recurrent unit network(GRU)and convolutional neural network(CNN)is built,and an attention mechanism is introduced to further extend the attention mechanism,and the feasibility is verified through experiments.The main research in this paper includes the following:For the problem that the PM2.5 concentration change process is a non-linear and non-smooth time series,a GRU network model is built to handle the prediction time series application for prediction.Then conventional prediction methods such as back propagation network(BP),CNN,recurrent neural network(RNN)and long short-term memory neural network(LSTM)were selected for comparison experiments.From the above models in PM2.5 concentration prediction,the root mean square error(RMSE),mean absolute error(MAE)and the coefficient of determination(R~2)of the GRU network were obtained to be better,respectively Therefore,the GRU network was chosen as the basis for a more in-depth study in this paper.Considering the problem of unstable model accuracy caused by the GRU model only considering the influence of the time factor and ignoring the influence of features,a CNN network that can perform feature extraction was introduced and a CNN-GRU prediction model was built.After experiments,the RMSE,MAE and R~2 of the CNN-GRU model are10.058,6.694 and 0.923 respectively,which are all better than the GRU model.Due to the problem of random weight assignment in the training of the CNN-GRU model,it was not possible to consider the factors influencing the PM2.5 concentration to play a different role in its variation,and an attention-based mechanism and CNN-GRU model were designed.The model divides the attention mechanisms into temporal attention,feature attention,and temporal+feature attention by classifying them according to the different dimensions of the role of attention mechanisms.The results of RMSE,MAE and R~2 for the three attention mechanisms were experimentally obtained as 9.467/9.433/9.535,6.372/6.168/6.406 and 0.932/0.933/0.931 respectively,which proved that the CNN-GRU model with the introduction of the three attention mechanisms all performed better than the CNN-GRU model.