Air Pollution Modelling and Forecasting in Hamilton Using Data-Driven Methods

Abstract

The purpose of this research is to provide an extensive evaluation of neural network models for the prediction and the simulation of some key air pollutants in Hamilton, Ontario, Canada. Hamilton experiences one of Canada's highest air pollution exposures because of the dual problem associated with continuing industrial emission and gradual increase of traffic related emissions along with the transboundary air pollutions from heavily industrialized neighboring north-eastern and mid-western US cities. These factors combined with meteorology, cause large degradation of Hamilton's air quality. Hence an appropriate and robust method is of most importance in order to get an early notification of the future air quality situation. Data driven methods such as neural networks (NNs) are becoming very popular due to their inherent capability to capture the complex non-linear relationships between pollutants, climatic and other non-climatic variables such as traffic variables, emission factors, etc. This study investigates dynamic neural networks, namely time lagged feed-forward neural network (TLFN), Bayesian neural network (BNN) and recurrent neural network (RNN) for short term forecasting. The results are being compared with the benchmark static multilayer perceptron (MLP) models. The analysis shows that TLFN model with its time delay memory and RNN with its adaptive memory has outperformed the static MLP models in ground level ozone (O_3) forecasting for up to 12 hours ahead. Furthermore the model developed using the annual database is able to map the variations in the seasonal concentrations. On the other hand, MLP model was quite competitive for nitrogen dioxide (NO_2) prediction when compared to the dynamic NN based models. The study further assesses the ability of the neural network models to generate pollutant concentrations at sites where sampling has not been done. Using these neural network models, data values were generated for total suspended particulate (TSP) and inhalable particulates (PM_10) concentrations. The obtained results show promising potential. Although there were under-predictions and over-predictions on some occasions, the neural network models, in general were able to generate the missing information and to obtain air quality situation in the study area.