Robust predictive control of lambda in internal combustion engines using neural networks

Stoichiometric air-to-fuel ratio (lambda) control plays a significant role on the performance of three way catalysts in the reduction of exhaust pollutants of Internal Combustion Engines (ICEs). The classic controllers, such as PI systems, could not result in robust control of lambda against exogenous disturbances and modeling uncertainties. Therefore, a Model Predictive Control (MPC) system is designed for robust control of lambda. As an accurate and control oriented model, a mean value model of a Spark Ignition (SI) engine is developed to generate simulation data of the engine's subsystems. Based on the simulation data, two neural networks models of the engine are generated. The identified Multi-Layer Perceptron (MLP) neural network model yields small verification error compared with that of the adaptive Radial Base Function (RBF) neural network model. Consequently, based on the MLP engine's model, the MPC system is performed through a nonlinear constrained optimization within gradient descent algorithm. The performance of the MPC system is compared with that of a first order Sliding Mode Control (SMC) system. According to simulation results, the tracking accuracy of lambda by the MPC system is close to that of the SMC system. However, the MPC system results in considerably smoother injected fuel signal.