A comparative study of feed forward neural networks and radial basis neural networks for modeling Tokamak fusion process

This work is aimed at simulating the neural network based state space models of the Tokamak fusion reactor. Two different types of neural networks have been used in this study to form state space neural networks, namely feedforward neural networks (FFNN) and radial basis neural networks (RBNN). The work presents analysis of FFNN and RBNN based state space models developed for Tokamak reactors. It has been found that the developed neural network state space models are computationally more efficient and equally accurate when compared to the standard state space models. However, initially some time investment is required to train the neural networks. The predictive quality of both FFNN and RBNN has been found to be similar. FFNN are preferred over the RBNN because of their overall less computational load. In general the application of neural networks resulted in time savings up to 95%. This saving in time is a function of number of states, inputs and outputs present in the original state space model.