A new structure identification scheme for ANFIS and its application for the simulation of virtual air pollution monitoring stations in urban areas

Parameter and structure identifications are necessary in any modelling which aims to achieve a generalised model. Although ANFIS (Adaptive Network-based Fuzzy Inference System) employs well-known parameter-identification techniques, it needs to structure identification techniques for the determination of an optimum number of fuzzy rules and the selection of significant input variables from among the candidate input variables. In this study, a new structure identification scheme is developed and introduced, which is simultaneously capable of the selection of significant input variables and the determination of an optimum number of rules. This new structure identification was joined to ANFIS, and this joined modelling framework was applied to the simulation of virtual air-pollution monitoring stations in Berlin. In this study, 18 virtual particulate matter stations were simulated using the particulate matter data of some of the current stations. In other words, the particulate matter monitoring network of Berlin has been intensified. The evaluation of simulated virtual stations shows that, although the uncertainty of daily particulate matter measurement is about 10 percent, the simulated virtual stations can estimate the mean daily particulate matter with less than 10 percent of error. Mean absolute error and root mean square error of the simulations are less than 2.4 and 3.4µg/m3, respectively. The correlation coefficient of the simulation results was more than 0.94. In addition, the range of mean bias error is between −1.0 and 0.5µg/m3, and the range of factor of exceedance is between −14.8 and 10.8 percent. It means that the simulated virtual stations have a small bias. These results demonstrated the appropriate performance of the joined new structure identification scheme and ANFIS for development of a virtual air pollution monitoring network.