fr | en
Laboratoire Angevin de Recherche en Ingénierie des Systèmes

Séparés par des virgules

Séminaire LARIS - Rayane CHMAYCEM11h | Polytech Angers | Salle du conseil 101 | 62 Notre Dame du Lac - 49000 Angers

"Graph Neural Networks for Urban Air Quality Forecasting: From Static to Dynamic and Multi-Graph Models"

Nitrogen dioxide (NO₂) is a major urban pollutant, driven mainly by road traffic and transported by the wind. Forecasting it at monitoring stations is a spatiotemporal problem: stations are scattered irregularly across the city, and their concentrations depend both on their own history and on what happens at neighbouring stations. Graph neural networks are well suited to this setting, as they represent the monitoring network as a graph whose nodes are stations and whose edges describe how stations influence each other. Node attributes can include hourly NO₂ concentrations, meteorological variables (temperature, humidity, wind speed and direction) and traffic measurements, while edge attributes can encode the distance between two stations, their alignment with the wind, or the similarity of their traffic conditions.

This talk shows how the choice of graph structure shapes forecasting performance. We first consider static GNNs, in which the graph structure and edge weights are fixed once and remain the same at every time step; here the graphs are built from inter-station distances and combined with feature-specific encoders and sparse graph transformers. We then move to dynamic GNNs, in which the edge weights, and possibly the connections themselves, change over time to reflect evolving conditions; we introduce physics-informed graphs whose edge weights evolve hourly according to wind direction and speed, so that information flows along the actual transport of pollutants. Finally, we discuss ongoing work on multi-graph models that combine a wind graph and a traffic-similarity graph through a learned fusion mechanism. These approaches are evaluated on hourly NO₂ data from Madrid and from the Paris dataset built during this thesis.

Rayane CHMAYCEM  est doctorante en 3ème année de thèse en co-tuelle au sein de l'équipe Information, Signal, Image et Sciences du Vivant (ISISV). Elle est encadrée par Nisrine JRAD au LARIS et Ahmad DIAB (LIU - Liban) et co-encadrée par Zaher MEHRI (LIU - Liban)

Scroll