Research On One Step And Multi-step Prediction Of PM2.5 Concentration

In recent years,with the rapid economic development of various countries in the world,the continuous improvement of social productivity,and the continuous enrichment of material and social life,the uncoordinated development of the economy and ecological environment has become increasingly prominent.The problem of air pollution is becoming more and more serious,and haze weather frequently occurs in many cities around the world.The appearance of haze weather not only brings many adverse effects on people’s normal life,work and study,but also aggravates the harm to people’s health and the sustainable development of society.Fine particles such as PM2.5 in air pollutants enter the human body through breathing,which is an important hazard to the human body.Scientifically predicting PM2.5 allows people to do protective work in advance and minimize the impact on human health.Scientific and reasonable prediction of PM2.5 has practical significance.This thesis collected hourly pollutant data for Chongqing from 2015 to 2017,and performed a variety of preprocessing on the data set,such as filling missing values and normalizing data.At the same time,the correlation analysis of various pollutants contained in the data set was also carried out.When predicting the PM2.5 concentration,a single-step duration prediction model CNN-TCN was first established.By establishing the spatio-temporal feature matrix of PM2.5 and using CNN multiple convolutions to extract the spatial feature data of pollutants between multiple stations.At the same time,PM2.5 concentration data is also time series data.By fusing the data features on the spatial and temporal levels,TCN uses dilated convolutions with different expansion coefficients to extract the data features,and finally outputs the predicted value through the fully connected layer.In the multi-step duration prediction of PM2.5 concentration,a sequence model Seq2Seq-CNN based on dual attention mechanism was proposed.The model can predict the data of different input duration through the sequence-to-sequence model structure,and at the same time it can also obtain the forecast data of different output duration.Seq2Seq-CNN,like CNN-TCN,uses CNN to extract features at the spatial level.In addition,it also uses feature attention to extract features from different pollutants in the site,and uses the time attention mechanism to generate different context vectors at different output times,so as to obtain more accurate PM2.5 concentration prediction value at the decoding part of the sequence.In view of the characteristics of serial data that can be input with different durations,this thesis also compares the prediction effects of data with different input durations.In this thesis,the prediction system of PM2.5 concentration in Chongqing is also established by using the prediction model of seq2seq-CNN based on dual attention,The system obtains the pollutant concentration data of each site in Chongqing in real time and forecasts after processing.Finally,the system vividly displays the historical data and prediction data of the site through the ECharts and Baidu map.