While methods to measure air pollution have become increasingly refined, prediction remains a challenge despite the development of sophisticated vehicular exhaust dispersion models. This is largely due to the complexity, non-linearity and unknown distributional qualities of air pollution data. In response, there is growing interest in using data-driven machine-learning techniques, such as artificial neural networks (ANN) to model air quality data. The appeal of ANNs is that they are capable of modelling highly non-linear functions and can be trained to accurately generalise from a new independent data set. ANNs are also good at detecting the underlying pattern masked by noisy factors in a complex, highly disaggregate, system. Despite the potential, ANN-based approaches have largely been applied to the problem of predicting regional or city-wide pollution. Relatively few applications have focused on roadside exposures. The current paper reports on the development and application of ANN-based methods to address the problem of temporally disaggregate-level prediction of particulates near a busy intersection in Sydney, Australia. Following details of the data collection required, the paper explains the rationale for the ANN structure used for this application and compares it to other modelling approaches before drawing conclusions on the merits of the approach. (a) For the covering entry of this conference, please see ITRD abstract no. E216058.
Abstract