Wind turbines play a major role in the transformation from a fossil fuel based energy production to a more sustainable production of energy. Total-cost-of-ownership is an important parameter when investors decide in which energy technology they should place their capital. Modern wind turbines are controlled by pitching the blades and by controlling the electro-magnetic torque of the generator, thus slowing the rotation of the blades. Improved control of wind turbines, leading to reduced fatigue loads, can be exploited by using less materials in the construction of the wind turbine or by reducing the need for maintenance of the wind turbine. Either way, better total-cost-of-ownership for wind turbine operators can be achieved by improved control of the wind turbines. Wind turbine control can be improved in two ways, by improving the model on which the controller bases its design or by improving the actual control algorithm. Both possibilities have been investigated in this thesis. The level of modeling detail has been expanded as dynamic in ow has been incorporated into the control design model where state-of-the-art controllers usually assume quasi-steady aerodynamics. Floating wind turbines have been suggested as an alternative to ground-fixed wind turbines as they can be placed at water depths usually thought outside the realm of wind turbine placement. The special challenges posed by controlling a floating wind turbine have been addressed in this thesis. Model predictive control (MPC) has been the foundation on which the control algorithms have been build. Three controllers are presented in the thesis. The first is based on four different linear model predictive controllers where appropriate switching conditions determine which controller is active. Constraint handling of actuator states such as pitch angle, pitch rate and pitch acceleration is the primary focus of this controller. The wind turbine is a highly nonlinear plant and a gain scheduling or relinearizing model predictive controller forms the next step to improve performance compared to a linear controller. Finally, a nonlinear model predictive controller has been devised and tested under simplified conditions. At present, the nonlinear model predictive controller is however not expected to be an realistic option for real world application as the computation burden is to heavy to achieve real-time performance. This thesis is comprised of a collection scientific papers dealing with the various topics presented in this summary.