Neural network inference measurements applied to the pebble bed modular reactor

Abstract

Inference measurements with time-delayed feed-forward neural networks facilitates the inference of unknown variables from known variables in non-linear dynamic systems. This is based only on the mapping data of the known variable and variable to be inferred. For successful inference, several constraints have to be overcome. This is, the neural network should have the correct topology, the training data set characteristics must have inherent attributes to ensure generalisation and the training algorithm must be capable of finding an acceptable local minimum on the error surface. At present, the neural network topology is based on trial and error, while the generalisation capability of the trained neural network is tested by using test and validation sets. Due to the lack of design methods for the topology of neural networks and the need for independent testing and validation, this thesis endeavours to develop a generalised method to find the optimum topology for accurate inference measurements. The aim is further to develop a method for judging the training set that could lead to generalisation without using test sets or validation sets. For this to be done, the training algorithm should succeed in finding a small enough local minimum on the error surface. The developed methods are applied to a simulated model of the pebble bed modular reactor (PBMR).