Investigating the applicability of spatiotemporal transformer networks for air quality forecasting: A case study in the Greater Helsinki region

Abstract

This thesis investigates the applicability of transformer networks in spatiotemporal time-series forecasting of air quality. With air pollution becoming an increasingly critical issue, accurate forecasting is essential for mitigation efforts. Air pollution data is typically organized both spatially and temporally, and this study aims to demonstrate that a spatiotemporal transformer architecture might be able to effectively incorporate both dimensions of information to generate more accurate results than other state-of-art models.

This study is a single case study, and it applies design research methods to develop and demonstrate the effectiveness of a new model architecture. Air pollutant data from the region of Greater Helsinki is used to train a spatiotemporal transformer model. This model is then utilized to forecast unknown PM2.5 values together with other comparative models, and the results are analyzed to provide insights on the performance differences.

The findings suggest that the presented spatiotemporal transformer model provides more accurate forecasts than comparative models. This has multiple managerial implications for decision-makers in both the private and public sectors. The model can improve public health outcomes, reduce healthcare costs, enhance environmental sustainability, and improve data-driven decision-making. The thesis also provides theoretical contributions by demonstrating the forecasting performance of transformer-based models in spatiotemporal scenarios, and how to train spatiotemporal transformer models with lower computational costs.