Current Profile Control in EAST via Reinforcement-Learning-based Model Predictive Control
Z. Wang, S. Morosohk, S.T. Paruchuri, E. Schuster
65th Division of Plasma Physics (DPP) Annual Meeting of the American Physical Society (APS)
Denver, CO, USA, October 30 – November 3, 2023
The spatial distribution of the toroidal plasma current density, simply
referred to as current profile, is crucial for the realization of advanced
modes of operation in tokamaks. A control-oriented model, governed by
the Magnetic Diffusion Equation (MDE), is discretized and linearized to
make it suitable for Model Predictive Control (MPC) design while simultaneously
reducing the associated computational burden. This MPC algorithm generates
optimal control strategies that satisfy the physical constraints imposed
by available actuators within the Experimental Advanced Superconducting
Tokamak (EAST), such as neutral beam injection and lower hybrid wave sources.
Furthermore, this work introduces an innovative methodological enhancement
by proposing a Reinforcement-Learning-based Model Predictive Control (RLMPC)
approach. The role of reinforcement learning is to counterbalance the
model uncertainties in the simplified control-oriented model employed
within the MPC strategy. Swift convergence of the tunable parameters within
the RLMPC is facilitated through the use of a second-order Least Square
Temporal Difference Q-learning (LSTDQ) algorithm. Simulation studies based
on COTSIM show that the proposed RLMPC adeptly tracks a desired profile
evolution, demonstrating robust resilience in the face of disturbances.
*Supported by the US DOE under DE-SC0010537.