Investigating a deep learning approach to real-time air quality prediction and visualisation on UK highways

Akinosho, Taofeek

Abstract

The construction of intercity highways by the United Kingdom (UK) government has resulted in a progressive increase in vehicle emissions and pollution from noise, dust, and vibrations amid growing concerns about air pollution. Existing roadside pollution monitoring devices have faced limitations due to their fixed locations, limited sensitivity, and inability to capture the full spatial variability, which can result in less accurate measurements of transient and fine-scale pollutants like nitrogen oxides and particulate matter. Reports on regional highways across the country are based on a limited number of fixed monitoring stations that are sometimes located far from the highway. These periodic and coarse-grained measurements cause inefficient highway air quality reporting, leading to inaccurate air quality forecasts. Multi-target neural network is a type of machine learning algorithm that offers the advantage of simultaneously predicting multiple pollutants, enhancing predictive accuracy and efficiency by capturing complex interdependencies among various air quality parameters. The potentials of this and similar multi-target prediction techniques are yet to be fully exploited in the air quality space due to the unavailability of the right data set. To address these limitations, this doctoral thesis proposes and implements a framework which adopts cutting-edge digital technologies such as Internet of Things, Big Data and Deep Learning for a more efficient way of capturing and forecasting traffic related air pollution (TRAP). The empirical component of the study involves a detailed comparative analysis of advanced predictive models, incorporating an enriched dataset that includes road elevation, vehicle emission factors, and background maps, alongside traditional traffic flow, weather, and pollution data. The research adopts a multi-target regression approach to forecast concentrations of NO2, PM2.5, and PM10 across multiple time steps. Various models were tested, with Fastai's tabular model, Prophet's time-series model, and scikit-learn's multioutput regressor being central to the experimentation. The Fastai model demonstrated superior performance, evidenced by its Root-Mean Square Error (RMSE) scores for each pollutant. Statistical analysis using the Friedman and Wilcoxon tests confirmed the Fastai model's significance, further supported by an algorithmic audit that identified key features contributing to the model's predictive power. This doctoral thesis not only advances the methodology for air quality monitoring and forecasting along highways but also lays the groundwork for future research aimed at refining air quality assessment practices and enhancing environmental health standards.

Citation

Akinosho, T. Investigating a deep learning approach to real-time air quality prediction and visualisation on UK highways. (Thesis). University of the West of England. https://uwe-repository.worktribe.com/output/11414285

• Thesis Type: Thesis
• Deposit Date: Nov 6, 2023
• Publicly Available Date: Jul 25, 2024
• Public URL: https://uwe-repository.worktribe.com/output/11414285
• Award Date: Jul 25, 2024