Control-Oriented Modelling of a Wankel Rotary Engine: A Synthesis Approach of State Space and Neural Networks 2020-01-0253
The use of Wankel rotary engines as a range extender has been recognised as an appealing method to enhance the performance of Hybrid Electric Vehicles (HEV). They are effective alternatives to conventional reciprocating piston engines due to their considerable merits such as lightness, compactness, and higher power-to-weight ratio. However, further improvements on Wankel engines in terms of fuel economy and emissions are still needed. The objective of this work is to provide an engine modelling methodology that is particularly suitable for the theoretical studies on Wankel engine dynamics and new control development.
In this paper, a control-oriented model is developed for a 225CS Wankel rotary engine produced by Advanced Innovative Engineering (UK) Ltd. Through a synthesis grey-box approach that combines State Space (SS) and artificial Neural Networks (NN), a model is derived by leveraging both first-principle knowledge and engine test data. We first re-investigate the classical physics-based Mean Value Engine Model (MVEM). It contains differential equations mixed with empirical equations, which are inherently nonlinear and coupled. Moreover, the rotary configuration of Wankel engine implies distinctive dynamics which the MVEM may not be able to cover. Therefore, we derive a SS formulation which introduces a compact control-oriented structure with limited computational demand. It can effectively deal with a Multi-Input, Multi-Output (MIMO) system and avoid the cumbersome structure of the MVEM. On the other hand, via nonlinear system identification techniques, we compare the three different NN structures that are suitable for engine modelling using time-series test data: 1) Multi-Layer Perceptron (MLP) models; 2) the Hammerstein-Wiener (HW) models; 3) Nonlinear AutoRegressive with eXogenous inputs (NARX) models. These black-box methods achieve higher accuracy than the SS model and do not require any a priori knowledge of the engine dynamics. The resulting NN can perform as a high-fidelity engine simulator with satisfactory accuracy.
Anthony Siming Chen, Giovanni Vorraro, Matthew Turner, Reza Islam, Guido Herrmann, Stuart Burgess, Chris Brace, James Turner, Nathan Bailey
University of Bristol, University of Bath, University of Manchester, Advanced Innovative Engineering (UK) Ltd.