Air Pollution Observation Station Clustering Based On Spatial And Temporal Data

With the rapid development of economy and the continuous improvement of people's living standard, air pollution problem has become more and more serious, especially in some large or medium-size cities. It is critical to trace the diffusion path of the polutant and locate the source so that we can prevent and control pollution efficiently.Conventional air pollution source tracing methods are based on pollutant components or diffusion models, which trace pollution sources in large area, without partitioning the cities into subnets. However, in large cities,the reason of air pollution is unclear and complicated. For example, there could be various pollution sources located in different areas, which makes it difficult in identifying all the pollution sources directly through limited amount of observation stations.To tackle this problem, we partition the Beijing city into subnetworks. According to the data from all the 35 pollution monitoring station in Beijing, we extract the feature vectors, based on which we partition the area into subnetworks by K-Means and Modularity models.In our early work, we modeled the connections of the pollution sources by Word Activation Force (WAF) model. With the affinity matrix obtained via WAF modeling, we conducted clustering of the air pollution stations. Compared to K-means clustering method, which is based on locations, the proposed approach showed several advantages in air pollution network clustering. However, the WAF model has limits in the heterogeneous complex data processing.Recently, deep learning provides new solutions for the feature extracting of heterogeneous complex data. Considering that air pollution data has complex components, various source and too many influencing factors, we apply deep belief networks (DBNs) to reduce dimensions of original data and exploring data relevance. Clustering stations based on features from DBNs, all the stations in the same subnet are more likely influenced by the same pollution source, which helps a lot in tracing the air pollution diffusion path.In this thesis, we also provide a novel way to trace pollution sources according to sequential clustering results.