Evaluating Spatiotemporal Graph Neural Network Architectural Design for Urban Air Pollution Forecasting

Air pollution inflicts severe damage to the human body and mind. Effective environmental policymaking can increase life expectancy in cities, informed by interpretable insights from air pollution forecasting. This work evaluates STGNN architectural design for the task of forecasting PM2.5 concentrations in units of PPM.Firstly, the dataset was produced by merging an air quality sensor network dataset collected by the Urban Air Project with long-term climatology data recorded by the Climate Knowledge Portal. Then, 45 STGNN architectures were trained on this dataset under identical hyperparameter and hardware configuration, for 3 trials each. These architectures differed in positional encoding fusion strategy, graph convolutional network, recurrent neural network, and presence of bidimensional attention mechanism. Architectures employing additive or concatenative positional encoding with attention, using DCN with LSTM or GRU, yielded the lowest relative forecasting error, with MAE below 20 PPM. Overall, both positional encoding and attention significantly improved predictive accuracy by increasing the quality of spatiotemporal representations, with the following exception. Models employing concatenative positional encoding, RNN, and attention displayed extremely high relative forecasting error, with MAE above 30 PPM. While STGNNs are capable of accurately forecasting air pollution, careful consideration of architectural design is paramount to ensure that models are well-calibrated and well-structured for the task. This work demonstrates that STGNNs with representational enhancements are effective for urban air pollution forecasting, supporting useful environmental policymaking to increase life expectancy for humanity.

Author

• Alexander, Moses Blessing

Language

• en

Keyword

• pollution
• spatiotemporal
• urban
• forecasting
• recurrent neural network
• graph convolutional network

Date created

• 2025

Type of Work

• Dissertation

License

• All rights reserved

Rights statement

• In Copyright

GW Unit

• Computer Science

Degree

• D.Engr.

Advisor

• Raghavan, Vijay

Committee Member(s)

• Focacci, Aaron
• Etemadi, Amir

Persistent URL

•  https://scholarspace.library.gwu.edu/etd/n583xv82j