Estimation of research reactor core parameters using cascade feed forward artificial neural networks

The pattern of the core reload program is very important for an optimize use of research reactors. Reactor safety issues and economic efficiency should be considered during pattern studies. In order to find the best core pattern for a research reactor, its reloading program should be solved as a multi-objective and constrained optimization problem. If considered objective functions of the optimization problem can be estimated in very short time, the optimal fuel reloading pattern can be used effectively. In this research a very fast estimation system for suggested core parameters has been developed using cascade feed-forward type of artificial neural networks (ANNs). Four main core parameters are suggested to optimize reactor core adequately. And also to get larger thermal fluxes in the desired irradiation box, a new flexible method was selected. A Software package has been developed to prepare and reform required data for ANNs training. The gradient descent method with momentum weight/bias learning rule has been used to train ANNs. To get the best conditions for considered ANNs training a vast study has been performed. It includes the effects of variation of hidden neurons, hidden layers, activation functions, learning and momentum coefficients, and also the number of training data sets on the training and simulation results. Some experimental convergence criteria are used to study them. A comparison selection rule has been used to adjust desirable conditions. Final training and simulation results show that developed ANNs can be trained and estimate suggested core parameters of research reactors very quickly. It improves effectively pattern optimization process of core reload program.