Many automotive systems such as engines have manufacturing tolerances or change over time. This limits the performance of controllers tuned for the nominal case. A robust controller can not always overcome this performance gap. Against this background, in this work, we propose a self-tuning control strategy for an engine air path model obtained from data of a real engine and show its benefits setting. The self-tuning control consists of an online parameter estimation algorithm for polynomial non-linear autoregressive with exogenous input (PNARX) models and a nonlinear model predictive controller (NMPC) implemented by the continuation/generalized minimum residual (C/GMRES) algorithm. In a first step design of experiments (DOE) is utilized to identify a PNARX model offline from measurements performed on an engine test bed. A tracking NMPC is designed for this model and applied in simulation on the identified model. The control performance is assessed for the case of a wrong initial guess. It is shown that the resulting performance gap can be overcome by the online parameter estimation of a k-step prediction model with directional forgetting. An improved closed loop control performance of the air path model confirms the approach.