Physics-based machine learning for predicting urban air pollution using decadal time series data

Abstract

Air quality significantly impacts public health and well-being. Effective air quality policies depend on accurate pollution predictions, even in countries with low pollution levels like Norway. Although LSTM networks are well-suited for time series data, they require large datasets and are computationally expensive, limiting their use for physical process data like air pollution dynamics. Physics-Based Machine Learning (PBML) integrates physical laws into models, offering more accurate and interpretable predictions for air pollution. In this study, we compare PBML, LSTM, and Linear Regression Models (LRM) to identify significant predictors of air pollution using daily traffic, weather, and air pollution data between 2009–2018 from three major Norwegian cities. Our findings demonstrate that PBML outperforms both LSTM and LRM in predicting air pollution levels. This paper contributes to the expanding literature on PBML by offering more precise air pollution predictions on hyperlocal scales, informed by local traffic and weather conditions, ultimately supporting better data-driven policy decisions.