• Energy Research

Predicting energy consumption has become a critical issue for energy-intensive industrial contexts. A significant contribution to their overall energy load is due to the Heating Ventilation and Air Conditioning (HVAC) systems. This work, therefore, aims to validate the applicability of a probabilistic graphical approach, the Bayesian Network, in predicting the HVAC systems’ energy consumption. As a data-driven approach, it is compared with more common AI-based models like Support Vector Machine, Artificial Neural Networks and Random Forest. The graphical approach ensures a better interpretation of the main factors determining the energy consumption and the relationships underlying these dependences. After an initial contextualisation and an analysis of the state of the art, the design methodology of a Bayesian network is investigated in detail, deepening in the various solutions for each step and evaluating their performance through the application on two industrial case studies. The results show that Bayesian networks, despite not always providing the best results, are a valid solution, trading off between simplicity, flexibility, and performance. Moreover, the possibility to provide a physical interpretation of the results is one of its main strengths. The critical aspect encountered, instead, is the need for discretisation, which strongly influences the quality of the results.