Source term estimation: variational method versus machine learning applied to urban air pollution

Roman Lopez-Ferber (1) , Sylvain Leirens (1) , Didier Georges (2)

1 DSYS - Département Systèmes
2 UGA - Université Grenoble Alpes

Source detection is a field of study gaining interest due to environmental concerns about air quality in populated areas. We developed a machine learning framework inspired by previous works on road traffic estimation, and compared it to a classical variational method under a unidimensional and stationary problem. We tested source reconstruction with datasets coming from 12 and 50 sensors with and without noise. Noise was set to follow a gaussian law with a dependent variance from the maximum measured value of a concentration profile. Both methods are reasonably robust to noise. The results reveal that the Neural Network used here, a multilayer perceptron, performs very well compared to the classical 3D-Var method.

Mots clés: Source Detection; Source Term Estimation; variational methods; 3D-Var; air pollution; neural network; advection-diffusion

Domaines: Traitement du signal et de l'image [eess.SP]; Machine Learning [stat.ML]; Sciences de l'environnement

Roman Lopez-Ferber, Sylvain Leirens, Didier Georges. Source term estimation: variational method versus machine learning applied to urban air pollution. CSC 2022 - IFAC Workshop on Control for Smart Cities, Jun 2022, Sozopol (virtual), Bulgaria. ⟨cea-03716399⟩