A new approach in improvement of mean value models for spark ignition engines using neural networks

• We present a highly accurate, real-time control-oriented model for SI engines.
• Incorporating neural nets into mean value models, we achieve a grey-box extension.
• Neuro-MVM is much more accurate than MVM, and is also more reliable than a mere NN.
• The model precisely predicts transient conditions, in a wide range of reliability.
• This study reaches high levels of accuracy in designing neural networks.

In this paper, a real-time, highly precise control-oriented model for multi-point fuel injection SI engines is presented. Basically, the first step in the majority of control procedures is modeling. This model is supposed to be capable of demonstrating the influences of control inputs on control outputs, in an accurate and real-time manner. Considering the increasing requirements for engines, presenting such a model (which is both real-time and accurate) for the effective development of the engine Electronic Control Unit (ECU) has been a challenging concern over the past years. Using artificial neural networks and incorporating them into conventional mean value models, this study is to present a real-time model with high accuracy and fidelity for spark ignition engines, as thorough as possible. Integrating neural networks (which are black-box models) into the approximate mathematical relations of engine (which are roughly white-box), we will achieve a grey-box extension in engine modeling. We name this model as Neuro-MVM. Not only is Neuro-MVM much more accurate than the prevalent mean value models, but it is also more robust (and of a higher reliability) compared to a mere black-box model. Moreover, it is appropriately real time. In this regard, first for data acquisition, a one-dimensional CFD model will be constructed and calibrated with the aid of the GT-Power software. It will be attempted to provide a comprehensive pattern table of the engine’s (desirable and undesirable) area of performance to train neural networks. Improving the customary fashions, this study attempts to reach high levels of accuracy in designing neural networks. As will be seen, the attained regressions will be line-like. More than anything, this is due to the principles and practices adopted in this survey. In the end, Neuro-MVM (which is implemented in SIMULINK) will be capable of precisely predicting engine’s (steady state and transient state) outputs such as emissions, produced torque, and engine speed in real time.