Research Of PM2.5 Prediction Based On Multi Phase Hybrid Model

With the rapid development of industrialization and urbanization,the air quality of many cities in China has been deteriorating.The serious air pollution events occur frequently and affect more and more widely.The harm to residents’ health and social economy is becoming more and more obvious.Fine particulate matter,especially PM2.5,is the main factor causing haze,reducing visibility and endangering human health.Therefore,it is of great significance to study and master the influencing factors and temporal and spatial distribution characteristics of regional atmospheric PM2.5 concentration,and to explore an effective regional atmospheric PM2.5 concentration prediction model for reducing the health risk of the public and providing scientific reference and decision-making basis for the implementation of regional air pollution control and joint prevention and control measures.In this paper,aiming at the problem of PM2.5 concentration prediction,the main research work is as follows:(1)A phase partition method is studied and a multi-phase PM2.5 concentration prediction model is establishedInfluenced by monsoon and heating emissions,PM2.5 concentration time series shows different behavior patterns in different periods.In order to better fit the time series of PM2.5 concentration in different periods and improve the prediction accuracy of PM2.5 prediction model,this paper proposes a multi-phase PM2.5 prediction method.Firstly,a time series clustering algorithm based on density peak warped kmeans(DPWKM)is proposed to divide the time periods.Secondly,a prediction model based on attention long short term memory network(AT-LSTM)is established in each sub-phase,which can extract long term macro dependence and improve the memory weight of key time steps.The results show that the PM2.5 prediction model based on multi-phase ATLSTM has the best performance in the data of different monitoring stations.(2)A PM2.5 concentration prediction method based on multi-phase hybird model is studiedPM2.5 concentration time series has complex chaotic and nonlinear characteristics,and it is difficult to fully grasp its characteristics with only one model.The hybird model can combine the advantages of different sub-models to achieve the purpose of complementary shortcomings.In order to further improve the accuracy and stability of the prediction model,a PM2.5 concentration prediction model based on multi-phase AT-LSTM-FCN is established by using AT-LSTM network to extract long-term macro dependence while using full convolutional network(FCN)based on compression excitation block to extract short-term local dependence.The experimental results show that the PM2.5 concentration prediction model based on multi-phase AT-LSTM-FCN can effectively track the complex fluctuations of PM2.5 concentration time series.(3)A PM2.5 concentration prediction method based on multi-phase and multi-task learning hybird model is studiedIn the same area,PM2.5 concentration time series of different monitoring stations have similar change patterns.In addition,due to the influence of diffusion,the pollutants between different stations also have a certain degree of interaction.Multitask learning method can mine and utilize the same characteristics hidden in different tasks,and ultimately improve the effect of each learning task.Therefore,this paper studies and establishes a combined prediction model based on multi-phase and multitask learning,using the bottom network of the model to learn the common evolution characteristics of PM2.5 concentration time series of different monitoring sites,and using the high-level network of the model to learn the exclusive characteristics of PM2.5 concentration time series of different monitoring sites respectively,to establish a multi-site PM2.5 concentration prediction model.