Aplicação de redes neurais para inferência de temperaturas históricas na usina nuclear Angra 1

Abstract

In this work, a solution proposal is presented for the absence of temperature history in Angra 1 Nuclear Power Plant, which wish to request its long-term operation. Equipment in the reactor containment building are susceptible to high temperatures, making its lifetime qualified to be different from that of project base. Ageing studies of this equipment are necessary to elaborate the environmental qualification program, document needed to request the long-term operation license. To elaborate this studies it is necessary a history of temperature which this equipment has been submitted over the years of operation. A usual problem found at the moment of ageing study is the absence of temperature history over the years in strategical positions to the ageing study. This work proposes a temperature inference methodology, simulating historical detectors through artifical neural networks. The methodology uses measurements of detectors installed since year 2015 together with measurements of the Sistema Integrado de Computadores de Angra. Two architectures of neural networks are trained, DNN and LSTM, with two goals, making the inference of temperature and to compare which architecture performs better in the database. Both architectures present satisfactory results at the end of the work, with LSTM presenting a result 25% better than DNN.