Data Driven Forecasting Models for Urban Air Pollution: MoreAir Case Study

Abstract

Artificial Intelligence has the potential to contribute to sustainable cities, life on land, and climate action. Specifically, data-driven AI models can analyze large, interconnected databases to develop joint environmental actions. Air quality plays a pivotal role in both climate action and the development of sustainable cities, but developing countries face challenges due to insufficient monitoring stations and limited access to air quality data sets. This study builds upon the MoreAir project, which established a low-cost air pollution monitoring system and provided the first air quality data set from Morocco. We first exploit and delve into the details of the obtained dataset. Subsequently, we conduct a multi-level comparison of data-driven forecasting models, specifically focusing on short-term forecasting of Particulate Matter concentrations. Four forecasting frameworks are explored, using different combinations of exogenous data and spatio-temporal information. Our findings highlight that Machine Learning models, particularly LightGBM and CatBoost, outperform other models. Overall, our study demonstrates that the inclusion of the spatial dimension along with the diverse exogenous features enhances the models’ predictive performance, and provides valuable insights.

Keywords

• Air pollution
• urban air pollution forecasting
• open datasets
• statistical models
• machine learning
• deep learning