Neural Network Based Fast-Running Engine Models for Control-Oriented Applications

A structured, semi-automatic method for reducing a high-fidelity engine model to a fast running one has been developed. The principle of this method rests on the fact that, under certain assumptions, the computationally expensive components of the simulation can be substituted with simpler ones. Thus, the computation speed increases substantially while the physical representation of the engine is retained to a large extent. The resulting model is not only suitable for fast running simulations, but also usable and updatable in later stages of the development process. The thrust of the method is that the calibration of the fast running components is achieved by use of automatically selected neural networks. Two illustrative examples demonstrate the methodology. The results show that the methodology achieves substantial increase in computation speed and satisfactory accuracy.