Research On Multi-Site PM2.5 Forecasting Method With Attention Mechanism

In recent years,air quality problems around the world,especially the haze,have made people pay more attention to environmental issues.PM2.5 is the main component of smog.In order to effectively respond to smog pollution,the need to grasp the development trend of PM2.5 concentration in advance has become stronger.However,in actual research and application,the influence of the surrounding environment and various data set factors brings certain challenges to PM2.5 prediction.How to efficiently and accurately predict PM2.5 concentration and provide guidance for air pollution protection has become the current research.Focus.Aiming at the background that most PM2.5 prediction models only consider the impact of time series and ignore the spatial factors and the correlation of pollutants,this paper proposes a multi-site prediction model of CNN-LSTM-attention based on deep learning.The data collected by the air quality monitoring station predicts the PM2.5concentration value of the target station in the next 1-5 hours,which improves the accuracy of the prediction to a certain extent.The main contents of this article are as follows:(1)Considering that PM2.5 is affected by changes in pollutant emissions,changes in time,and changes in meteorological factors,this article adds meteorological data to air pollutant data,and uses the short-term memory neural network(LSTM)model as the initial prediction model.At the same time,in order to increase the calculation weight of the key factors affecting PM2.5 at different points in time,an improved attention mechanism was introduced on the LSTM,and more attention was paid to sequence features that are more critical to PM2.5 impact,giving full play to the LSTM time.Memory advantage and ability to express features.(2)Considering that the change in PM2.5 concentration has a region-dependent characteristic,diffusion and transfer will occur over time,which affects the prediction of PM2.5 concentration in neighboring stations.This paper proposes a multi-site model prediction method based on the fixed proximity radius method.The characteristics of the time series are extracted while taking into account the influence of neighboring station information,and the diversity of input data is further expanded.By adopting the method of sharing parameters at multiple sites,the complexity of the network to be trained is not increased,and the risk of network overfitting is reduced.(3)In order to study the spatial correlation between PM2.5 and air pollutants,and fully exploit the spatial characteristics of propagation between sites,this paper introduces a convolutional neural network(CNN)to extract the spatial characteristics between data.Emphasizing the integration of time series relationships and spatial characteristics between sites makes the model’s features more diverse,gives full play to the advantages of each network,and greatly improves the efficiency of network training.In order to verify the effectiveness of the method proposed in this article,experiments were performed on each model constructed and comparative analysis was performed.The results show that the CNN-LSTM multi-site PM2.5 prediction model integrated with attention mechanism predicts the RMSE index of 0.267 in the next hour The MAE index is0.191,which is significantly better than other models.At the same time,the long-term prediction performance is at the optimal index compared with the rest of the models,indicating that the problem of predicting longer time intervals is also solved to a certain extent,and the effectiveness of the model proposed in this paper is fully verified.