Grey-box modeling architectures for rotational dynamic control in automotive engines

Real-time model-based control of complex automotive systems rely on developing high fidelity models which balance accuracy and computation load. These models can be developed by using grey-box modeling techniques. In this paper, a feedforward artificial neural networks model is combined with a physical engine model to form a grey-box model which predicts engine rotational dynamics. The grey-box model with different architectures is tested using a range of transient experimental data. In particular, serial and parallel error-based grey-box architectures are compared to determine the better architecture. The results demonstrate the grey-box model significantly enhance the prediction results from the engine physical model. The serial grey-box is found as the preferred architecture but the parallel grey-box can reach the same performance as that of the serial grey-box if the embedded physical model inside the grey-box has sufficient accuracy.