The development of a dense urban air pollution monitoring network

Abstract

The importance of air pollution monitoring networks in urban areas is well known because of their miscellaneous applications. At the beginning of the 1990s, Berlin had more than 40 particulate matter monitoring stations, whereas, by 2013, there were only 12 stations. In this study, a new and free–of–charge methodology for the densifying of the PM10 monitoring network of Berlin is presented. It endeavors to find the non–linear relationship between the hourly PM10 concentration of the still–operating PM10 monitoring stations and the shut–down stations by using the Artificial Neural Network (ANN), and, consequently, the results of the shut–down stations were simulated and re–constructed. However, input–variables selection is a pre–requisite for any ANN simulation, and hence a new fuzzy–heuristic input selection has been developed and joined to the ANN for the simulation. The hourly PM10 concentrations of the 20 shut–down stations were simulated and re–constructed. The mean error, bias and absolute error of the simulations were 27.7%, –0.03 (μg/m3), and 7.4 (μg/m3), respectively. Then, the simulated hourly PM10 concentration data were converted to a daily scale and the performance of ANN models which were developed for the simulation of the daily PM10 data were evaluated (correlation coefficient >0.94). These appropriate results imply the ability of the developed input selection technique to make the appropriate selection of the input variables, and it can be introduced as a new input variable selection for the ANN. In addition, a dense PM10 monitoring network was developed by the combination of both the re–constructed (20 stations) and the current (12 stations) stations. This dense monitoring network was applied in order to determine a reliable mean annual PM10 concentration in the different areas in Berlin in 2012.