Prediction Of PM_2.5 Concentration By Fusing Spatio-temporal Correlation

Nowadays,air pollution is an important issue that has a negative impact on human health and the environment.Excessive PM_2.5 concentration is one of the main reasons for air pollution,and accurate and effective prediction of PM_2.5 concentration is an important foundation for achieving air quality pollution prevention and control.Therefore,this article proposes a PM_2.5 concentration prediction model that integrates spatiotemporal correlation,using a combination of neural networks and attention mechanisms to improve the accuracy of the prediction.Firstly,the influencing factors and changing trends of PM_2.5 concentration were analyzed.This article analyzes the degree to which PM_2.5 concentration is influenced by various pollutants and meteorological characteristics by calculating the Pearson correlation coefficient between pollutant concentration and meteorological characteristics,as well as PM_2.5 concentration.The time dimension of year,month,and day measures were used to plot the trend of pollutant concentration changes and analyze the temporal evolution relationship of PM_2.5 concentration.The sequence similarity of multi-site PM_2.5concentration was analyzed from the spatial dimension,and the interaction characteristics between spatial sites with PM_2.5 concentration were found.Secondly,a feature selection based LSTM-ATT time dimension PM_2.5 concentration prediction model was established.Based on random forest,the importance ranking under feature coupling is obtained,and the optimal feature subset is selected by combining the correlation analysis results of random forest and Pearson.Based on the influence of changes in pollutant emissions,meteorological factors,and time on site PM_2.5,an improved prediction model LSTM-ATT is proposed by integrating attention mechanism to improve the accuracy of PM_2.5 concentration prediction.Once again,a PM_2.5 concentration prediction model integrating spatiotemporal correlation was proposed.Based on LSTM-ATT method,temporal features are extracted,and spatial interaction of PM_2.5 concentration is further analyzed.Based on CNN-LSTM-ATT,spatial features are studied,and multi site concentration is transformed into a supervised learning sequence.Multilayer convolution is designed to extract spatial features.Attention mechanism is used to allocate different weights of site features,and the site impact effect is learned through training.By combining spatiotemporal dimension prediction with fully connected layers,accurate prediction of air quality PM_2.5concentration was achieved by integrating spatiotemporal correlation.Finally,taking the pollutant concentration and meteorological data set of 12 stations in Beijing as the research object,improved the mean interpolation method to fill in the missing values,combined with Pearson correlation feature ranking and random forest importance ranking to obtain the feature subset,and predicted the PM_2.5 concentration in the next 500 and 1000 hours in the time dimension,spatial dimension and fusion model respectively,The effectiveness and accuracy of the PM_2.5 concentration prediction model integrated with spatiotemporal correlation in this project have been verified through comparison.