Using Neural Networks to Predict Core Parameters in a Boiling Water Reactor

The problem of optimizing refueling in a nuclear boiling water reactor is difficult since it concerns combinatorial optimization and it is NP-Complete. In order to solve this problem, many techniques have been applied, ranging from expert systems to genetic algorithms. In most of these procedures, nuclear reactor simulators are used, which require a longer computation time, to evaluate the goodness of the proposed solutions. As the processes are iterative, many evaluations with the simulator are necessary, and this makes the process extremely slow. In this paper, the use of trained neural networks (NNs) is proposed as an alternative to the simulator, and the results of the NN training are shown in order to predict some variables of interest in the optimization, such as the effective multiplication factor and some thermal limits, related to safety aspects. Finally, a study about the effect of modifying several NN parameters is shown.