Offset free tracking in Model Predictive Control requires estimation of unmeasured disturbances or the inclusion of an integrator. An algorithm for estimation of an unknown disturbance based on adaptive estimation with time varying forgetting is introduced and benchmarked against the classical disturbance modelling approach, where the system description is augmented by a disturbance state. The time varying forgetting renders the new approach less sensitive to the nature of the disturbance. By simulation we demonstrate that this algorithm is advantageous in case of infrequent step disturbances of any magnitude.