A hybrid model-based and learning-based approach for the plasma magnetic control in DEMO

Abstract In this paper we propose a new hybrid approach, based on a deep reinforcement learning technique and a model-based technique, for the solution of the control problems regarding the magnetic confinement of a plasma in the DEMO tokamak. Reinforcement learning agents are used together with classical model-based controllers to perform the magnetic confinement of the plasma, i.e., to control the position, the shape and the current of the plasma. This hybrid approach allows us to simplify the training procedure of the data-driven control policy and to improve the performance of the model-based solutions. The performance of the proposed approach is shown in numerical simulations by evaluating the vertical stabilization capability and the error in tracking references on the plasma current and shape.