Prediction of particulate and dissolved 137Cs concentrations in the Rhône River (France) using machine learning - IRSN - Institut de radioprotection et de sûreté nucléaire Accéder directement au contenu
Communication Dans Un Congrès Année :

Prediction of particulate and dissolved 137Cs concentrations in the Rhône River (France) using machine learning


The worldwide presence of artificial radioactivity in rivers due to the nuclear installations and the various historical fallouts is proven for many years. For sanitary reasons, it is necessary to know and predict the levels of radioactivity in these rivers, both in dissolved and particulate fractions. Thus, for several decades, the radioactivity of rivers is measured by monitoring networks. Despite they have been improved with the advancements on technology and knowledge on radionuclides, these networks remain subject to logistical constraints which will result in gaps in the monitoring records. To fill these gaps, mechanistic or semi-empirical models are possible solutions, but they require parameters that are difficult to obtain and that must be often estimated, generating uncertainties in the predicted concentrations. In the recent decades, machine learning approaches were extensively developed allowing new modelling approaches based on the numerical analysis of large datasets acquired by monitoring networks (data-driven models). Because, these approaches are no longer depending on theoretical or empirical parameters, this study illustrates how they can be used to fill gaps in empirical series of dissolved and particulate 137Cs concentrations. Modeling was applied to such concentrations acquired since several decades by the monitoring of the Rhône River in the south of France which is a highly nuclearized river affected by the atmospheric fallouts of the nuclear tests and the Chernobyl accident and by the controlled liquid radioactive discharges of several nuclear power plants and other nuclear facilities. Different Machine learning approaches are applied to link and extrapolate these 137Cs concentration series with monitoring series of hydrological parameters such as the water discharge and the suspended sediment flux and the 137Cs liquid discharges of the Reprocessing Center of Marcoule which is the main cesium-emitting nuclear facility in the Rhône River. Algorithms such as random-forest, gradient-boosting or neural network have been tested and compared. The results show that the dissolved and particulate 137Cs concentrations are accurately modeled directly from the hydrological parameters without it being necessary to specify the liquid discharges of the nuclear facilities. The comparison with a semi-empirical model based on solid/liquid fractionation shows that this data-driven model approach can be more operational in some cases. Its main limitation concerns specific and/or extremely rare events which are not included in the dataset and which were, consequently, not correctly learned by the numerical learning procedure. Finally, in the context of rivers monitoring, it is also showed how these approaches can be used as an inverse calculation method to assess the liquid 137Cs discharges of a facility from the river monitoring data.
Tuesday_Session 1c_Lepage.pdf (1.46 Mo) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)
Licence : CC BY NC ND - Paternité - Pas d'utilisation commerciale - Pas de modification

Dates et versions

irsn-04065984 , version 1 (12-04-2023)


  • HAL Id : irsn-04065984 , version 1


Hugo Lepage, Valerie Nicoulaud Gouin, Patrick Boyer. Prediction of particulate and dissolved 137Cs concentrations in the Rhône River (France) using machine learning. 5th International Conference on Radioecology & Environmental Radioactivity, Sep 2022, Oslo, Norway. ⟨irsn-04065984⟩
3 Consultations
1 Téléchargements


Gmail Facebook Twitter LinkedIn More