Deep learning-based multi-target regression for traffic-related air pollution forecasting

Abstract

Copyright © 2023 The Author(s). Traffic-related air pollution (TRAP) remains one of the main contributors to urban pollution and its impact on climate change cannot be overemphasised. Experts in developed countries strive to make optimal use of traffic and air quality data to gain valuable insights into its effect on public health. Over the years, the research community has developed advanced methods of forecasting traffic-related pollution using several machine learning methods albeit with persistent accuracy and insufficient data challenges. Despite the potentials of emerging techniques such as multi-target deep neural network to achieve optimal solutions, they are yet to be fully exploited in the air quality space due to their complexity and unavailability of the right training data. It is to this end that this study investigates the impact of integrating an updated data set including road elevation, vehicle emissions factor and background maps with traffic flow, weather and pollution data on TRAP forecasting. To explore the robustness and adaptability of our methodology, the study was carried out in one major city (London), one smaller city (Newport) and one large town (Chepstow) in the United Kingdom. The forecasting task was modelled as a multi-target regression problem and experiments were carried out to predict NO2, PM2.5 and PM10 concentrations over multiple timesteps. Fastai’s tabular model was used alongside prophet’s time-series model and scikit-learn’s multioutputregressor for experimentation with fastai recording the overall best performance. Statistical tests run using Friedman and Wilcoxon test also revealed the significance of the fastai model with a p-values < 0.05. Finally, a model explanation tool was then used to reveal the most and least influential features from the newly curated data set. Results showed traffic count and speed were part of the most contributing features. This result demonstrates the impact of these and other introduced features on TRAP forecasting and will serve as a foundation for related studies.