Data analysis and identification of zone-based neural network models of an automotive Heating, Ventilation and Air Conditioning system

A Heating, Ventilation and Air Conditioning (HVAC) system is the object of study of this thesis. The main objective is the development of suitable model to reproduce the system behavior, in collaboration with the company "DENSO Thermal Systems". The final model will be the plant of a control system, but the controller development is not treated in the present thesis. It is not a traditional HVAC unit, but a next generation one that finds application in the context of a luxury vehicle. The novelty is the introduction of a decoupled zone management, whose consequence is a significantly higher number of operating modes, compared with a conventional HVAC. In particular, this system characteristic results in a higher flexibility in managing the unit in the vehicle, but concurrently, in an increase in modelcomplexity. For the already cited reasons, instead of classical white box approaches, that in this case are too difficult to deal with, a black box approach is chosen, to model the system from experimental measurements collected on a test bench. The present thesis illustrates the two main phases characterizing the activities carried out. The first is about data exploration and analysis, which was essential to reveal intrinsic aspects of the system, using techniques such as: descriptive statistics, correlation analysis, Principal Component Analysis and k-means Clustering. After this first analysis, the exploration of suitable model architectures was the focus of the second phase, but, before that, techniques like feature importance and Shapley analysis had a crucial role in generating ideas for model simplification and focusing on the most valuable features. At first, several models based on traditional regression and on kernel-regression were designed. Subsequently, modeling techniques based on Neural Networks constituted the focus of the work. A single Multi-Input Multi-Output (MIMO) network was trained, in order to predict all the twelve outputs concurrently and using all the available inputs. Then, a zone-based architecture was employed, creating four MIMO networks. In conclusion, a Multi-Branch approach, resulting from the fusion of the previous two architectures, gave the best results in terms of performances and robustness. Overall, the nonlinear dynamic of the analyzed HVAC unit was not fully captured by linear and semi-linear models. In fact, the thesis demonstrates that Neural Networks are the most suitable and reliable approaches for this type of applications.