Robust neural predictive control of normalized air-to-fuel ratio in internal combustion engines

In this paper, a neural predictive controller (NPC) is designed to control emission pollutants of ground vehicles. The proposed controller is designed based on robust structure of model predictive control (MPC) system in order to control normalized air-to-fuel ratio (lambda) within ±1% of the stoichiometric value. As an accurate and control oriented model of engine, a mean value engine model (MVEM) of a spark ignition engine is developed to generate simulation data. Engine model identification is preformed through an off-line multi-layer Perceptron neural network (MLPN) which is trained by gradient descent back propagation algorithm. In the controller, an on-line MLPN is designed to generate optimum control action signals of the closed loop system. The performance of the new controller is compared with the performance of a standard MPC system which is using constrained minimization of an introduced cost function through Gradient Descent (GD) algorithm. According to the simulation results, the calculation time cost of the NPC is significantly smaller than the standard model predictive systems. Moreover, the proposed controller is satisfactorily robust to engine time varying dynamics and unstructured uncertainties.