Performance assessment of recurrent neural network-based engine models for the estimation of combustion indexes

The environmental and political context is leading the automotive world to evolve towards a more sustainable future, not only from an environmental but also from an economic and social point of view. This is why research in this field is becoming increasingly in-depth, particularly in the development of technologies and control systems for internal combustion engines. In this context, artificial neural networks (ANN) offer a very promising approach for modelling and controlling internal combustion engines. In particular, the new technologies and requirements of this transition pose a great challenge for engineers of the future. Thus, this thesis aims to study a methodology to design neural networks that can predict, given appropriate inputs, combustion indices accurately. The first objective is therefore to define all the main parameters that characterise the architecture of a neural network and optimise them for the problem studied. The network's task is to predict as accurately as possible the value of MFB50 (50% of fuel burnt mass). The networks were trained from steady-state data and then tested on dynamic profiles, precisely to begin to tackle the above problem. Therefore, an attempt was made to identify a robust methodology to define a network structure of good accuracy. Secondly, the aim is to evaluate the behaviour of these networks by providing the same architectures with different input variables. In particular, we want to study the influence of VVT variables and thus observe whether networks trained using these input variables perform better or worse in understanding test profiles.