A mobile dose prediction system based on artificial neural networks for NPP emergencies with radioactive material releases

• Autor(a) principal: PEREIRA, Claudio M. N. A.
• Data de Publicação: 2017
• Outros Autores: SCHIRRU, Roberto, GOMES, Kelcio J., CUNHA, José Luiz, Instituto de Engenharia Nuclear (IEN), Universidade Federal do Rio de Janeiro - PEN/COPPE/UFRJ
• Tipo de documento: Artigo de conferência
• Idioma: eng
• Título da fonte: Repositório Institucional do IEN
• Texto Completo: http://carpedien.ien.gov.br:8080/handle/ien/2153
• topic: Dose prediction; Atmospheric dispersion of radionuclide; Mobile

This work presents the approach of a mobile dose prediction system for NPP emergencies with nuclear material release. The objective is to provide extra support to field teams decisions when plant information systems are not available. However, predicting doses due to atmospheric dispersion of radionuclide generally requires execution of complex and computationally intensive physical models. In order to allow such predictions to be made by using limited computational resources such as mobile phones, it is proposed the use of artificial neural networks (ANN) previously trained (offline) with data generated by precise simulations using the NPP atmospheric dispersion system. Typical situations for each postulated accident and respective source terms, as well as a wide range of meteorological conditions have been considered. As a first step, several ANN architectures have been investigated in order to evaluate their ability for dose prediction in hypothetical scenarios in the vicinity of CNAAA Brazilian NPP, in Angra dos Reis, Brazil. As a result, good generalization and a correlation coefficient of 0.99 was achieved for a validation data set (untrained patterns). Then, selected ANNs have been coded in Java programming language to run as an Android application aimed to plot the spatial dose distribution into a map.In this paper, the general architecture of the proposed system is described; numerical results and comparisons between investigated ANN architectures are discussed; performance and limitations of running the Application into a commercial mobile phone are evaluated and possible improvements and future works are pointed.