Faire avancer la sûreté nucléaire

La Recherchev2


Data assimilation for the management of environmental and radiological data in case of nuclear accident.



Email :

Mot de passe :

N. Pérot
Actes du congrès ECORAD, 3-7 sept 2001, Aix en Provence, France
Radioprotection - Colloques, volume 37, C1-1139/ C1-1145.

Type de document > *Article de revue

Mots clés > modélisation en radioécologie, bases de données, situation accidentelle

Unité de recherche > IRSN/DEI/SECRE/LME

Auteurs >

Date de publication > 01/07/2002


The "Data Assimilation for the Management of Environmental and Radiological Data in Case of Nuclear Accidents" project managed by the IRSN/DPRE started in 2001 for at least three years. It is based on the IRSN skills in the field of data analysis and the modelling of radionuclide transfers in the environment more precisely along the food chain pathways. The main part of the project consists in defining methods to fit radioecological models with incoming measured values when predictions do not correspond to such measured values. In such a case, the radioecological parameters of the model or some poorly assumed initial conditions are supposed to be responsible and have to be modified. In order to update the modeling, the "parameters" of the model have to be adjusted considering the measured values. This problem of data assimilation can be represented by a nonlinear program (or optimization problem), where the relations between the parameters and their definition domains constitute the set of constraints; then, the cost function to minimize is the difference between predictions and measured values. The experts can control the process of optimization by adding some well chosen constraints expressing their understanding of the phenomena considered in the studied case.

Problems of non-linear programs are known to be very difficult to solve and the calculation time is rarely reasonable deterministic methods, which scan all the research space. In this context, it is proposed to treat this problem by using a combination of a deterministic method such as Constraints Satisfaction Problems (CSP) and a stochastic method with a genetic algorithm to reach best performances.