Outdoor air pollution (AP) is a serious public threat which has been linked to severe respiratory and cardiovascular illnesses, and premature deaths especially among those residing in highly urbanised cities. As such, there is a need to develop early-warning and risk management tools to alleviate its effects. The main objective of this research is to develop AP forecasting models based on Artificial Neural Networks (ANNs) according to an identified model-building protocol from existing related works. Plain, hybrid and ensemble ANN model architectures were developed to estimate the temporal and spatiotemporal variability of hourly NO2 levels in several locations in the Greater London area. Wavelet decomposition was integrated with Multilayer Perceptron (MLP) and Long Short-term Memory (LSTM) models to address the issue of high variability of AP data and improve the estimation of peak AP levels. Block-splitting and crossvalidation procedures have been adapted to validate the models based on Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Willmott's index of agreement (IA). The results of the proposed models present better performance than those from the benchmark models. For instance, the proposed wavelet-based hybrid approach provided 39.15% and 28.58% reductions in RMSE and MAE indices, respectively, on the performance of the benchmark MLP model results for the temporal forecasting of NO2 levels. The same approach reduced the RMSE and MAE indices of the benchmark LSTM model results by 12.45% and 20.08%, respectively, for the spatiotemporal estimation of NO2 levels in one site at Central London. The proposed hybrid deep learning approach offers great potential to be operational in providing air pollution forecasts in areas without a reliable database. The model-building protocol adapted in this thesis can also be applied to studies using measurements from other sites.
|Date of Award||15 Sep 2020|
- University Of Strathclyde
|Sponsors||University of Strathclyde|
|Supervisor||Ben Richard Hughes (Supervisor) & Paul Gerard Tuohy (Supervisor)|