PM2.5 Concentration Data Treatment Method Based On Machine Learning Model

Air quality has received widespread attention in recent years and is critical to protecting the health of residents.Among them,fine particulate matter(PM2.5)is closely related to the atmospheric environment and human life.Affected by factors such as equipment performance and deployment cost,the number of PM2.5 monitoring stations in cities is limited and cannot provide fine-grained PM2.5 concentrations.How to infer,predict and calibrate the PM2.5 concentration in areas without monitoring stations based on existing monitoring stations is of great significance for residents’ health and urban planning and construction.In this paper,data from air quality monitoring stations,remote sensing,and meteorology were obtained,and the study area was divided into grids of the same size.Research.The specific content and results are as follows:(1)In view of the problem that the existing PM2.5 concentration inference model lacks the ability to establish a multi-order correlation coefficient matrix based on dynamic spatiotemporal characteristics,which leads to the inability to accurately infer the concentration of highly polluted PM2.5,this paper proposes an Multi-order Spatial-Temporal Graph Convolutional Inference model(MOSTGCNInf).The model uses the graph convolution network to extract the feature relationship,and uses the attention mechanism to dynamically construct the attention coefficient matrix of multi-level nodes,and performs spatiotemporal feature fusion to improve the PM2.5 concentration inference effect.MOSTGCNInf is mainly composed of four independent modules: the Attention module,which uses a multi-order attention mechanism to effectively capture the dynamic correlation in the sample data and build the attention coefficient matrix;the Multi-order graph convolution module,which convolves the sample data and Multi-order attention coefficient matrix and fusion;Fully-connected module,which transforms the output through linear space to generate PM2.5 concentration prediction probability distribution;Selftraining optimization module,which uses pseudo-label technology to iteratively infer PM2.5 concentration PM2.5 concentrations of unknown nodes.The experimental results show that MOSTGCNInf can significantly improve the Accuracy and F1 value of PM2.5concentration inference.(2)In view of the problem that the existing prediction model does not consider the correlation between PM2.5 concentration and adjacent areas,which leads to the inability to accurately predict the concentration of high-pollution PM2.5,this paper proposes a Balanced Social LSTM Prediction model(BSLPre).BSLPre is mainly composed of two modules: On the one hand,in order to capture the spatiotemporal correlation of PM2.5concentrations between adjacent sites,this paper develops a social LSTM model,which can effectively capture PM2.5 concentration changes at adjacent sites;On the other hand,considering the unbalanced effects caused by pollution levels in different regions,this paper designs a new B-MSE loss function to assign different weights to monitoring stations.The experimental results show that BSLPre can significantly improve the MAE and RMSE values predicted by PM2.5 concentration.(3)Aiming at the inaccurate measurement results of low-cost portable air quality monitoring sensors,a Category-Based Calibration Model(CCM)is proposed,which uses machine learning algorithms to process such portable sensors.Compared to traditional models that tend to learn a single regression model,CCM covers multiple regression models according to pollutant concentration categories and builds a more accurate mapping from sensor readings to reference values.In addition,CCM introduces two fault tolerance modules: classification error and sample fault tolerance.The former mitigates the impact of misclassification on lumped categories,and the latter improves the robustness of individual regression models.Experimental results show that CCM outperforms the baseline model in both MAE and SMAPE evaluation metrics.The above work effectively solves the problem of PM2.5 concentration inference,prediction and calibration under high pollution conditions,and significantly improves evaluation indicators such as accuracy and F1 value.