Research On Air Quality Prediction Method Based On Spatial-temporal Data

Air pollution is a problem that cannot be ignored in urban development,and pollutants such as PM2.5 that affect air quality have a serious impact on human health,so it is important to make fine-scale predictions for air quality to combat air pollution.Since the air quality data are typical spatio-temporal series data,this thesis uses a deep learning approach to model and predict spatio-temporal series data to build a set of air quality prediction models.The research of this thesis is as follows.Firstly,to address the problem that sparse spatial attribute data cannot fully reflect spatial characteristics,this thesis proposes an improved spatial interpolation method based on the third law of geology,which interpolates for areas with high environmental similarity by introducing a concept of environmental similarity,and then interpolates for the remaining unknown study area using the Kriging method after the first round of interpolation.Using this interpolation method can effectively avoid the problem of interpolation failure of the kriging interpolation method when the known data are sparse.Secondly,by analyzing the spatio-temporal periodicity characteristics of air quality data and combining with the air quality interpolation model described in the previous thesis,for the problem of spatio-temporal correlation of data with spatio-temporal characteristics,this thesis adopts a convolutional long and short-term memory network as the base module of the model,which can fully capture the spatio-temporal correlation characteristics of the data,and encode the spatio-temporal features by adopting an encoder-decoder structure,and through the encoder decoder structure to decode the encoded features and finally obtain a feature with high spatio-temporal correlation.At the same time,since the network is a variant of the long and short-term memory network,it can effectively avoid the problem of gradient disappearance and gradient explosion that occurs in recurrent neural networks for long-term prediction,and can effectively capture the dynamic evolution characteristics of spatio-temporal data,i.e.,learn and predict the periodic and trend characteristics of air quality data.Finally,the effectiveness of the spatial interpolation method proposed in this thesis and the codec-based neural network model is verified through comparison experiments.Experiments are designed to compare the errors of the spatial interpolation method proposed in this thesis with traditional spatial interpolation methods such as the inverse square of distance method and kriging method to verify the accuracy of the improved interpolation method based on the third law of geodesics.Another experiments are designed to verify the effectiveness of the spatial interpolation method by comparing the prediction results after using the spatial interpolation method with those without the spatial interpolation method.The air quality prediction model proposed in this thesis is compared with moving average autoregressive method,support vector regression method,and long and short-term memory network to verify the accuracy of the model proposed in this thesis.Single-step prediction and multi-step prediction are performed to verify the accuracy of the model proposed in this thesis in multi-step prediction.