Author(s)

Sheen Mclean Cabaneros, John Kaiser Calautit & Ben Hughes

Abstract

Roadside air pollution is a major issue due to its adverse effects on human health and the environment. This highlights the need for parsimonious and robust forecasting tools that help vulnerable members of the public reduce their exposure to harmful air pollutants. Recent results in air pollution forecasting applications include the use of hybrid models based on non-linear autoregressive artificial neural networks (ANN) with exogenous multi-variable inputs (NARX) and wavelet decomposition techniques. However, attempts employing both methods into one hybrid modelling system have not been widely made. Hence, this work further investigates the utilisation of wavelet-based NARX-ANN models in the shortand long-term prediction of hourly NO₂ concentration levels. The models were trained using emissions and meteorological data collected from a busy roadside site in Central London, United Kingdom from January to December 2015. A discrete wavelet transformation technique was then implemented to address the highly variable characteristic of the collected NO₂ concentration data. Overall results exhibit the superiority of the wavelet-based NARX-ANN models improving the accuracy of the benchmark NARX-ANN model results by up to 6% in terms of explained variance. The proposed models also provide fairly accurate long-term forecasts, explaining 68–76% of the variance of actual NO₂ data. In conclusion, the findings of this study demonstrate the high potential of wavelet-based NARX-ANN models as alternative tools in short- and long-term forecasting of air pollutants in urban environments.

Keywords

artificial neural networks, air pollution, air pollution forecasting, NARX, wavelet transform