Approcci model-free and data-driven per la stabilizzazione verticale del plasma nei tokamak

This doctoral dissertation examines the development of vertical stabilization of the tokamak plasma to advance magnetic confinement in nuclear fusion research. It investigates the potential of developing magnetic controllers by combining traditional control engineering techniques with Artificial Intelligence. This thesis provides a basis for using model-free and data-driven approaches as an alternative to the commonly used model-based vertical stabilization controllers. Model-based controllers have shown significant promise in the vertical stabilization of the tokamak plasma; however, their effectiveness can be limited by the complexity and uncertainty of plasma dynamics, potential model mismatch, and computational requirements. The work in this thesis addresses these limitations by developing control strategies that guarantee the required level of performance without relying on the knowledge of a plant model to improve the robustness of the overall plasma magnetic control system. To this aim, two model-free and data-driven approaches have been developed for vertical stabilization: the first one relies on the Extremum Seeking algorithm to achieve stabilization, while the second one is based on Reinforcement Learning. Most of the proposed controllers have been tested in simulation by considering the~ITER plasma as a case study. Specifically, a model-free Extremum Seeking algorithm for stabilization has been studied and deployed in the Vertical Stabilization system. To assess the robustness of the proposed approach, linear and nonlinear simulations were performed considering~ITER and~TCV tokamaks. In addition, automata for the adaptation of the real-time control gain adaptation and neural networks were included in the Extremum Seeking-based controller to enhance the robustness and generalization property of the~ITER Vertical Stabilization even in the presence of significant model uncertainties. Indeed, the Extremum Seeking algorithm calls for a Kalman filter to compute the required Lyapunov function, making the corresponding approach only quasi-model-independent. To resolve this issue, neural networks have been trained to estimate the plasma unstable dynamic and replace the Kalman filter in the control scheme. With the use of neural networks, the controller becomes completely model-free and the operative space of the Vertical Stabilization system is also enlarged by making it possible to stabilize plasma equilibria that were not stabilized by the set-up based on a single Kalman filter. Finally, Reinforcement Learning algorithms were considered to deploy an intelligent agent as a Vertical Stabilization system for the magnetic confinement of tokamak plasmas. A tabular Q-learning algorithm was first developed for the Vertical Stabilization of the~EAST tokamaks. It was followed by a Deep Deterministic Policy Gradient algorithm, which exploits an actor-critic setup based on deep neural networks to approximate the optimal behavior of the agent. The latter was implemented considering the entire magnetic control system of~ITER as an environment.