PM2.5 forecasting in a large city: Comparison of three methods

There is an increasing awareness for the toxic effects produced by the inhalation of fine particles present in the air. It is important then to provide precise information to the population about the concentrations of this pollutant expected for the incoming hours. We present here a study about the capability of three types of methods for PM2.5 forecasting one day in advance: a multilayer neural network, a linear algorithm and a clustering algorithm. Input variables are past concentrations measured in four monitoring stations and actual and forecasted meteorological information. Outputs are the maxima of the 24 h moving average of PM2.5 concentrations for the following day at the site of the monitoring stations. By training with data from the three previous years, we are able to generate results for the fall–winter period for each year from 2004 to 2007. Although the three methods may be used as operational tools, the clustering algorithm seems more accurate in detecting high concentration situations.