Equilibrium parameters recovery for experimental data in ASDEX Upgrade elongated plasmas

An application of neural network models to experimental ASDEX Upgrade data generated during the operational phases of the experiment is presented. The performances of the proposed models are compared to those obtained by using the realtime version of the function parameterization (FP) technique currently implemented on the plasma control computer of the machine. Both neural network and FP techniques generate a model of the underlying physical system by means of different learning procedures carried out on the same database of simulated experiments. In response to practical problems encountered in the analysis of the real plasma discharges, a number of modifications to the basic neural network fitting have been carried out. The outcome of the comparison between the methods encourages the study of the alternative neural technique, which appears to be significantly faster than FP in the final online implementation