Model Based Predictive Control (MBPC) is a well-known control technique especially in chemical industries and recently in aeronautics. Generalized Predictive Control (GPC) is an algorithm of MBPC family. Although existence of adaptive version of GPC, problems such as need for accurate modelling, linearization and online adapta-tion to system variations are still questions. In this paper a new control method is proposed named Linear Neural Generalized Predictive Control (LNGPC). It uses a combination of Linear Neural Networks (LNN) and GPC. LNN is used as parametric identifier that makes both linearization and pa-rameter extraction. An Algorithm is proposed for online learning of system Input/ Output bounds and making corresponding weight scaling. This ensures the stability of learning process even with variable system bounds. Online batch learning is used in LNN identification. An interpretation to the LNN weights is proposed to get system parameters as a discrete Transfer Function (TF). This given known order TF is then passed to a standard GPC controller.
Automatic Flight Control Systems (AFCS) often faces many problems such as un-modeled dynamics and fast parametric variations. LNGPC is tested with a realistic longitudinal rotorcraft model in a terrain-following application. The simulations show good results in terms of stability and adaptation comparing to other control schemes.