Control-oriented Modeling of the q-profile Dynamic Response in Preparation for Advanced Scenario Control in KSTAR

E. Schuster, S. Morosohk, T. Rafiq, H.-S. Kim, S.-H. Hahn, Y.-M. Jeon, M.D. Boyer, S. A. Sabbagh, M.L. Walker, D.A. Humphreys

KSTAR Conference

Seoul, South Korea, February 20-22, 2019

Active scenario control has the potential of playing a critical role in enabling access to high-performance modes of operation in KSTAR. As critical components such as the Motional Stark Effect diagnostic become available in real time, the development of a plasma response model arises as the next step towards achieving the scientific understanding of the current profile response dynamics in KSTAR that is necessary for control design. The high dimensionality of the current-profile control problem in tokamaks, along with the strong nonlinear coupling between magnetic and kinetic profiles, motivates the use of model-based control synthesis that can accommodate this complexity through embedding the known physics within the design. The distinctive characteristic of the modeling approach for current profile control followed in DIII-D and other long-pulse superconducting devices like EAST is the combination of widely accepted first-principle laws with data-driven correlations obtained from physical observations, which leads to a so-called first-principle-driven (FPD) model based on partial differential equations that capture the high dimensionality and the nonlinear magnetic-kinetic coupling of the system dynamics. Machine learning technique can play an important role in obtaining the data-driven correlations [1] that are necessary to close the control-oriented response model. FPD modeling has the potential of retaining the nonlinear dynamics of the plasma in the control-oriented model, which is critical for acceptable feedforward+feedback profile control design based on both offline and real-time model-based optimization. Moreover, FPD modeling provides the freedom of arbitrarily handling the trade-off between the simplicity of the model and both its predictive accuracy and its range of validity during the model reduction process prior to control synthesis (simulation-oriented model  (model reduction)  synthesis-oriented model). Present and planned work toward extending this technique to KSTAR will be discussed, including the development of a database to fully characterize the plasma state, which is defined by the equilibrium magnetic configuration, kinetic (density and temperature) profiles, magnetic (safety factor) profile, inductive and non-inductive current density profiles.

[1] S. Morosohk et al., “Towards the Development of a Neural-Network Version of NUBEAM for Scenario Optimization and Control in KSTAR,” this conference.