With this approach you do linearization followed by discretization which is generally less accurate than discretization followed by linearization. When you are integrating the linear approximation instead of the original nonlinear dynamics, the error in the linearization will grow with the sample interval. In this situation, the assumption is that linearized MPC is good enough over a time horizon T_s N, which if true means that the error incurred due to linearization over the horizon T_s only should be relatively small, so you might be fine.
Small update: I am now using a dual-loop controller, the inner loop controls the turn rate. This gives already a much better performance:
Better disturbance rejection, better damping, much easier to tune and works over a much larger range of wind speeds. Just a dual PI controller. So for this part of the control problem NMPC is not really needed, but there are other parts, in particular the winch control where it still might be useful.
Nice, cascade control is used for motion control in most industrial robot control systems also. Are you sure you need an integrator in both controllers? It’s typically only required in the inner controller. The outer controller will automatically benefit from the integration in the inner controller without having an integrator itself.