Research On Long-term And Short-term Air Quality Prediction Model Based On Temporal And Spatial Fusion

In recent years,the global air pollution situation is increasingly serious,frequent heavy air pollution events to the world’s economic development and human health sounded the alarm.High precision air quality prediction is the scientific basis and important means to provide environmental heavy pollution warning and guide air pollution prevention and control.Traditional deep learning models,due to their single structure,are not ideal in the face of complex data such as air quality.In order to more accurately fit the change trend of air quality in the future,this thesis carried out research from two aspects of air quality short-term prediction and long-term prediction.The primary contents are as follows:(1)Properly mastering the characteristics of the research data can effectively assist the reasonable construction of the later prediction model.In this thesis,Pearson correlation analysis and autocorrelation coefficient were used to conduct statistical analysis on the collected air quality data.The results show that air quality has obvious correlation characteristics in both time dimension and space dimension.(2)In this thesis,a prediction model of Spatio Temporal Air Quality Index(STAQI),which can capture complex spatio-temporal relationships at the same time,is applied to the short-term one-step prediction of Air Quality on the basis of retaining the geographical location distribution information of several stations.By improving the input data processing part of the gated recurrent unit network,the input data first flows through the graph convolutional network to learn the spatial features of the data,and then through the time sequence network to extract the time features between different time sequences.The whole STAQI model uses local component modeling to analyze the temporal characteristics of various pollutant concentrations in the site,uses global component modeling to model the spatial characteristics of air quality transmission in the surrounding sites,and finally connects the deep spatial and temporal representation learned by the components with appropriate weights to obtain the theoretically explicable model prediction value.The RMSE index value of STAQI model decreased by about 19%compared with the Gated Recurrent Unit model which had better prediction results among various comparison models.This indicates that the STAQI model combined with spatiotemporal modeling analysis is efficient and feasible for short-term air quality prediction.(3)In terms of time feature extraction,the traditional method stores the effective information obtained from model training in the hidden nodes and weights of each layer of the network,but the information of these limited memories is too small for long-term air quality prediction,which is not enough to obtain high-precision prediction results.In this thesis,a Time Series Memory Network(TSMN)model is applied to the long-term multi-step air quality prediction.TSMN model can be divided into two parts according to its structure,which are local memory component of time modeling and neighborhood component of space modeling.By introducing external memory into the network,the local memory component can improve the long-term memory ability of the model in the time dimension.Compared with CNN-LSTM model with better prediction effect,RMSE and MAE of TSMN model decreased by 5.4%,7.9% and R~2 increased by 3.6% respectively.By memorizing more useful information in time steps,the TSMN model can be supported to achieve high precision air quality prediction results in a long range.