Air Quality Inference Based On Spatio-temporal Big Data

With the rapid development of urbanization and the improvement of residents’ level,air quality has gradually become a hot issue concerned by residents and governments around the world.Air quality and related data can be collected by sensors deployed in different areas,and then the spatial and temporal fine-grained air quality distribution map can be obtained by using the inference algorithm,which is of great significance for urban air pollution protection and improvement of residents’ health and living standards.Therefore,how to accurately infer and forecast air quality at city scale has become an important problem to be solved urgently.The existing air quality inference methods mainly collect historical measurement data from air quality monitoring stations to infer air quality.However,the sparse air quality stations,incomplete data on relevant characteristics,various complex factors affecting air quality changes,and linear and nonlinear variations of air quality with location and time pose great challenges to accurate air quality inference.To address the problem of inferring air quality values in areas without monitoring stations using measurements collected from existing monitoring stations,a matrix decomposition-based approach is proposed to infer air quality by fusing information from low-rank structures,air quality measurements,and various features(e.g.,meteorology,traffic,road networks,land use,and points of interest).Unlike the existing research work that deals with feature recovery,feature extraction and air quality inference separately by reasoning about each,the three problems are effectively unified into a single model.In the model,the spatial and temporal feature matrices are decomposed into smaller factors to enable spatio-temporal implicit feature recovery and extraction.By sharing the matrix factors with the air quality matrix,the extracted spatial and temporal feature information is transferred to the air quality inference in order to improve its performance.The proposed method is evaluated based on a real data source obtained in Beijing.Experimental results show that the performance of the proposed method is improved by nearly an order of magnitude over that of the benchmark method.Aiming at the problem of using historical measurement data collected from target monitoring stations to predict future air quality values,a spatio-temporal deep learning based air quality prediction model was proposed.First,missing values are estimated by integrating improved inverse distance weighting(IDW)and improved autoregressive moving average(ARMA)models based on spatio-temporal view.Secondly,one-dimensional convolutional neural network(CNN)was used to capture the deep features of the geographical space between the stations,and long and short term memory network(LSTM)was used to extract the time dependence of the air quality series data.Finally,the prediction results are output through the full connection layer.The proposed model is evaluated based on real data sources obtained in Beijing.Experimental results show that the prediction performance of this model is better than that of the benchmark method.