Offset-free model predictive control for an active magnetic bearing system
Thapa Chhetri, Bhisma (2026)
Diplomityö
Thapa Chhetri, Bhisma
2026
School of Energy Systems, Konetekniikka
Kaikki oikeudet pidätetään.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi-fe20260803114817
https://urn.fi/URN:NBN:fi-fe20260803114817
Tiivistelmä
Active magnetic bearing systems are challenging to control because the rotor is open-loop unstable, the magnetic force is nonlinear, and the coil currents are limited by the actuator hardware. Model predictive control is useful for this type of system because it can handle multivariable dynamics and input constraints in the control problem. However, standard MPC can still leave a steady-state tracking offset when constant disturbances, modelling errors, or plant-model mismatch are present. This thesis studies OF-MPC for active magnetic bearing systems through simulation. The control structure is based on disturbance augmentation, Kalman-filter-based state and disturbance estimation, steady-state target correction, and constrained finite-horizon MPC. The OF-MPC formulation is first verified on a one-DOF AMB model by comparing it with standard MPC under constant disturbance and plant-model mismatch. The same control structure is then applied to a cone-shaped AMB with coupled axial and radial rotor motion. In the cone-shaped AMB simulations, the nonlinear plant is controlled using both linearized and nonlinear prediction models. The controller is tested for rotor lift-up from an initial displaced position and for short impact disturbances. The closed-loop response is assessed using tracking error, final offset, coil-current use, disturbance rejection, and computation time. The simulations show that OF-MPC reduces the steady-state error in the one-degree-of-freedom case and controls the rotor position in the cone-shaped AMB while keeping the coil currents within their specified limits. For the selected lift-up case, the linearized prediction model achieves the lowest RMS tracking error, while the nonlinear models produce smaller final offsets. The acados implementation requires less nonlinear MPC solution time than IPOPT.
