Sistema preditivo baseado em redes recorrentes para controle de reatores nucleares de pesquisa

Abstract

The control of nuclear reactors is done by systems that monitor several variables simultaneously, but only indicate the occurrence of an accident, without identifying its type or cause. The possibility of prediction of an accident based on the evolution of the current variables of the reactor can be a very useful tool to assist the work of the reactor operators, increasing their confidence, as well as the general availability of the plant. Echo State Networks (ESN) are Recurrent Neural Networks (RNNs) suitable for the formulation of predictive systems, with the advantage of using a training algorithm that is simpler and faster than conventional RNNs. To use Echo State Networks in predictive systems it is not necessary to specify the structure of the model, but some parameters of this network must be tuned for good performance. This work proposes the use of Artificial Intelligence to optimize ESN’s parameters, to improve performance. Particle Swarm Optimization (PSO), a method that uses so-called swarm intelligence, was used. The ESN optimized by the PSO was used to monitor the operation of a nuclear reactor (with real operating data), as well as to identify in real time if an accident is occurring, as well which accident is occurring. The results show that the network can identify and predict the state of the reactor, be it in normal or accident condition.