Forecasting powder dispersion in a complex environment using Artificial Neural Networks

Pierre Lauret (1) , Frederic Heymes (1) , Serge Forestier (2) , Laurent Aprin (1) , Alexis Pey (2) , Marcia Perrin (3)

1 LGEI - Laboratoire de Génie de l'Environnement Industriel
2 TÜV SÜD
3 Novartis Pharma AG

Atmospheric dispersion prediction skill is required for any industry processing hazardous material. This is a sensitive task since many parameters are involved: source term, atmospheric conditions, and local configuration. Behavior of dust dispersion is difficult because of the diameter scattering, agglomeration, sedimentation, range of densities... Furthermore, production sites may be located inside a complex environment such as urban areas, where accuracy of classical dispersion models is low. This paper aims to evaluate the efficiency of an Artificial Neural Networks (ANN) model to predict dust dispersion in an urban area without prior knowledge of the source term. The experimental database consists of 290 daily mean concentration measurements on a site located 500 m away from the emission source. The inputs are selected from meteorological data from a MeteoSwiss station located 4.5 km south. The training phase is done through early stopping application. ANN model selection is performed on the best coefficient of determination value. Model performance is evaluated using classical air quality criteria and shows good results. Nevertheless, ANN model tends to underestimate high concentrations while overestimating low concentrations. Results are included within acceptable range. Improvements can be achieved by adding information of the source term as an input for the ANN model.

Domaines: Mécanique des fluides [physics.class-ph]; Dynamique des Fluides [physics.flu-dyn]; Machine Learning [stat.ML]

Pierre Lauret, Frederic Heymes, Serge Forestier, Laurent Aprin, Alexis Pey, et al.. Forecasting powder dispersion in a complex environment using Artificial Neural Networks. Process Safety and Environmental Protection, 2017, 110, pp.71-76. ⟨10.1016/j.psep.2017.02.003⟩. ⟨hal-01962476⟩