On the use of artificial neural networks in loading pattern optimisation of advanced gas-cooled reactors

Artificial Neural Networks (ANNs) are applied to in-core fuel management optimisation of Advanced Gas-Cooled Reactors operated by British Energy in the United Kingdom to predict various parameters generated by the reactor core analysis code PANTHER. ANNs are biologically inspired computational models. Loading Pattern (LP) optimisation based on genetic algorithms (GAs) is being used in nuclear reactor fuel management studies, which demands substantial CPU times (at least order of weeks on a 866 MHz single processor PC) for multi-cycle problems. This is due to the assessment and qualification of large number of LPs required to solve an optimisation problem with PANTHER. This paper reports on the use of ANNs in predicting core physics parameters to improve the speed to obtain optimal candidate LPs. The construction and training of a number of ANNs to accelerate the optimisation process are described with the aim of using these networks as surrogate models. The supervised learning method has been used to carry out network training. We used three-layered feed-forward networks composed of one input, one hidden and one output layer with the backpropagation of error algorithm as the learning function. In addition the merits of using the scaled conjugate gradient learning algorithm has also been investigated. Several ways of using ANNs to accelerate optimisation are presented. Results have shown that ANNs recognise LPs that violate a radial power shape constraint and exclude (filter) those LPs from the search. We have demonstrated that ANNs can be used as accelerator algorithms within the GA. An attempt has been made to replace the PANTHER code with ANNs to perform a full multi-cycle optimisation for two AGR stations, which apply off-load and on-load refuelling. Results from these cases have shown orders of magnitude increases in speed. (author)