Modélisation de la pollution atmosphérique due au trafic maritime à Marseille

Modeling of atmospheric pollution due to ship traffic in Marseille

• Auteur / Autrice : Elliot Chevet
• Direction : Olivier Boiron, Fabien Anselmet
• Type : Projet de thèse
• Discipline(s) : Sciences pour l'ingénieur : spécialité Mécanique et Physique des Fluides
• Date : Inscription en doctorat le 01/10/2021
• Etablissement(s) : Aix-Marseille
• Ecole(s) doctorale(s) : Ecole Doctorale Sciences pour l'Ingénieur : Mécanique, Physique, Micro et Nanoélectronique
• Partenaire(s) de recherche : Laboratoire : IRPHE - Institut de Recherche sur les Phénomènes Hors Equilibre

Mots clés

Mots clés libres

Dispersion de polluants

Modélisation numérique

Pollution Maritime

Apprentissage automatique

Pollutant dispersion

Numerical modeling

Ship pollution

Machine Learning

Résumé

Le sujet de thèse porte sur la pollution atmosphérique des villes côtières méditerranéennes due au trafic maritime et, plus généralement, de toutes les grandes villes concernées par un trafic maritime intense. En utilisant le modèle de simulation de pointe pour le transport et la chimie des polluants atmosphériques, WRF/Chem, le comportement des émissions des navires en termes de concentrations de SO2 , NOX , PM 10 , PM 2.5 , O3, CO et CO2 sera estimé en fonction des variations saisonnières et du trafic, à échelle mésoscopique (1-2km). L'ensemble des données de prévision obtenu sera analysé à l'aide de techniques de Machine Learning pour tenter d'étendre l'analyse des principaux polluants (PM, O3, ...) en fonction de la météorologie. L'objectif général étant d'obtenir un outil de prévision fondé sur des données pour la prévision opérationnelle de la qualité de l'air.

Air pollution has been a persistent issue in Marseille and its surroundings for decades, with maritime traffic increasingly recognized as a key contributor to this public health concern. The port's proximity to the city center, combined with emissions from numerous ships at berth, including some of the world's largest cruise ships, has raised significant concerns and sparked extensive public debate. The pollutants of interest in this work, which include particulate matter (PM10, PM2.5), nitrogen dioxide (NO2), sulfur dioxide (SO2) and ozone (O3), are responsible for respiratory diseases such as airway irritation, cardiopulmonary and lung cancers. The PhD work presented here takes place in this context, with the first objective of setting up a numerical model to assess ships contribution to atmospheric pollution in Marseille. The model used to represent the many meteorological and chemical phenomena interacting in the atmosphere is the WRF-Chem model. To ensure an accurate representation of ship emissions, real-time traffic data from port authorities are used, in contrast to models based on annualized maritime traffic data. Marseille's unique geographical configuration, surrounded by valleys and bordered by the Mediterranean Sea, is influenced by specific phenomena that have a direct impact on air pollution in the region. A better understanding of these mechanisms constitutes a second objective of this research, as it would help to identify typical meteorological conditions that favor higher pollution levels during periods of intense maritime traffic. Providing daily atmospheric pollution forecasts is crucial for timely decision-making to mitigate the risk and informing the population directly affected by pollution. However, the computational cost of high-resolution simulations presents challenges for operational forecasting. The third objective of the PhD is addressing these chal- lenges by proposing a Machine Learning (ML) approach, supervised by numerical simulation results from the WRF-Chem model. Due to the high dimensionality of the data, traditional ML models face learning difficulties. To mitigate this, reduced- order modeling (ROM) techniques were applied to reduce data dimensions while retaining key information. Both Proper Orthogonal Decomposition (POD) and autoencoders were combined with a recurrent neural network (RNN) to predict PM2.5 concentrations. The main findings of this PhD research are: (i) ship pollution in Marseille The con- tribution of ships to air pollution in Marseille is estimated at 6 to 7% for PM2.5, 4% for PM10, and 18% for NO2; (ii) fine particles are transported inland while coarser particles remain near port areas due to gravitational settling; (iii) weak winds and sea-land breeze dynamics contribute to pollutant accumulation over the city; and (iv) the development of a multivariate POD-LSTM model significantly enhances long-term prediction accuracy compared to the univariate method.