Control-Oriented Simulations with Neural Network Based Semi-Physical Engine Models

摘要

This paper presents a novel integrated methodology for the conversion of high-fidelity engine models to fast running ones. The developed solution does not use the conventional system-level modeling approach, which creates a barrier between engine design and control system testing; rather it remains in the component-level domain, and proposes the substitution of the computationally expensive components of the simulation with computationally fast ones. The calibration of the fast running components is achieved through automatically trained neural networks. The thrust of the method is that the level of simplicity - and thus the execution speed - can be adapted to the specific application needs and computer power. In addition, the entire model conversion is made in a single integrated modeling environment, which promotes model sharing and bridges the gap between engine design and control-oriented applications.