---

The prerequisite in the majority of control processes is modeling. The model used to design a controller must be both accurate and real-time. Utilizing prevalent approaches of modeling, namely modeling based on (numerically) solving the equations governing the fluid in the combustion chamber, is too time-consuming and not suitable for a control purpose. This paper is to model combustion in an SI engine by means of neural networks and present an accurate and fast-response model for combustion. Obviously, any training procedure of neural networks does involve empirical data acquisition. On the other hand, engine testing is highly expensive, and testing data tables available (in industry) are not sufficient to train neural networks. In this paper, first with the aid of a CFD software, a one-dimensional model of an engine is constructed, and then calibrated using factual experimental data at hand. Afterwards, acquiring data required is performed via the validated CFD model. As a matter of fact, because of not having access to necessary experimental coefficients, calibration is an extremely complicated and time-consuming process. It will be attempted to accomplish and spell out the calibration of the engine model in the GT-Power software, in a scientific practice. After a brief survey on the methods employed in designing the neural networks, modeling of the combustion chamber will be stated. Eventually, the response of the constructed NN model will be compared to the results gained from the GT-Power software, and the great accuracy of the NN model will be indicated.

---

In this paper, an improved, real-time, highly accurate control-oriented style, named Neuro Mean Value Modeling, is presented for IC engine modeling. This model is a combination of neural networks and mean value model, and is able to overcome the shortcomings of both styles. In other words, taking advantage of both methods, this -box extension will be of more reliability than a mere black-box neural network, and also of more accuracy than roughly white-box mathematical relations of In this paper, the model is modified to become suitable for designing an engine controller. Thanks to the sophisticated methods applied (such as committee method, improved partitioning, and especially, simplifying neural networks’ tasks), neural networks of high accuracy with line-like regressions will be achieved. As will be seen, the model is precisely validated - and it is capable of accurately predicting the engine’s outputs (such as pollutant emissions, manifold pressure, knock probability, and engine speed) all in real time. In the end, the effect of engine control inputs on pollutant emissions and fuel consumption will be examined. The engine employed to establish the model is a ported fuel injection SI engine.