Development of an ANN–based air pollution forecasting system with explicit knowledge through sensitivity analysis

Abstract

Little attention is given to applying the artificial neural network (ANN) modeling technique to understand site–specific air pollution dispersion mechanisms, the order of importance of meteorological variables in determining concentrations as well as the important time scales that influence emission patterns. In this paper, we propose a methodology for extracting the key information from routinely–available meteorological parameters and the emission pattern of sources present throughout the year (e.g. traffic emissions) to build a reliable and physically–based ANN air pollution forecasting tool. The methodology is tested by modeling NO2 concentrations at a site near a major highway in Auckland, New Zealand. The basic model consists of an ANN model for predicting NO2 concentrations using eight predictor variables: wind speed, wind direction, solar radiation, temperature, relative humidity, as well as “hour of the day”, “day of the week” and “month of the year” representing the time variations in emissions according to their corresponding time scales. Of the three input optimization techniques explored in this study, namely a genetic algorithm, forward selection, and backward elimination, the genetic algorithm technique gave predictions resulting in the smallest mean absolute error. The nature of the internal nonlinear function of the trained genetically–optimized neural network model was then extracted based on the response of the model to perturbations to individual predictor variables through sensitivity analyses. A simplified model, based on the successive removal of the least significant meteorological predictor variables, was then developed until subsequent removal resulted in a significant decrease in model performance. The developed ANN model was found to outperform a linear regression model based on the same input parameters. The proposed approach illustrates how the ANN modeling technique can be used to identify the key meteorological variables required to adequately capture the temporal variability in air pollution concentrations for a specific scenario.