LINEAR SYSTEM IDENTIFICATION FOR PLASMA TRANSPORT

Abstract

The field of fusion control has been dominated by traditional physics-based dynamics models, but lately attention is been given to machine learning models such as neural networks. However, these models are often complicated and hard to analyze. In contrast, this project aims to capture the dynamics of the plasma transport in the DIII-D tokamak with simpler, linear models, which could be investigated under the methodology of system identification. This research investigates two main system identification models: state-space models of varying orders and autoregressive-moving average with exogenous terms models, ARMAX. It also uses tools such as standardization and principal component analysis to cope with the magnitude and complexity of the plasma data. Plasma shots, which are collected from the DIII-D tokamak in San Diego, are parsed into training and testing data to first train the aforementioned models and then to test them to evaluate their accuracy. In order to understand the performance of the models two key error measures are used: the mean squared error of the full plasma profiles, and the scatter plots of the delta profiles with correlation measures. The results indicate that the state-space models performed better in higher orders with the past profiles inputted to the model. The best state-space model is also compared to a selection of neural-net models and is found to be comparable in density and pressure predictions, while being better in safety factor and worse in temperature. Subsequently, the ARMAX models yielded a big difference between the two initial conditions deployed. The ‘optimal’ initial condition models yielded better results than both the models of ‘pseudo-inverse’ initial condition, and more importantly the neural-net models.