Improving Air Pollution Forecasting in Smart Cities using Clustering Techniques

Abstract

Air quality is a main concern for smart cities policies nowadays. This paper presents an approach for predicting Particulate Matter (PM) for air pollution using partitioning clustering techniques. Instead of predicting PM10 values for the entire dataset, better results were obtained when forecasting air pollution values for each discovered cluster. In the clustering stage, the k-means algorithm was applied, and four clusters were discovered. It could be noticed that two clusters corresponded to normal PM10 values, having the centroid values equal to 16, respectively 1, and the other two clusters stored high rates of pollution (with centroid values equal to 53, respectively 34). The forecasting results were more accurate when learning a cluster at a time with a specific classifier. After this step, several forecasting models were applied for each obtained cluster, and a conclusion that K-Nearest Neighbors and Neural Networks models had best performance in predicting the PM10 values is finally made.