Application of Data Analytics techniques for the analysis of building energy performance during operation : the case of Politecnico di Torino.

In 2019, building operations were responsible for about the 28% of the global CO2 emissions, taking into account not only the share directly due to the daily activity, but also the indirect part produced by the generation of power that supplies the building. The buildings consumption is strongly affected by the Energy Performance Gap, which is the deviation of the actual energy performance of the building with respect to the expected and designed one. As a consequence, the scope of improvement is relevant for the buildings and, in this context, an effective energy management has a key role. The aim of this thesis work is to provide a data analytics methodology whose results can be helpful to increase the knowledge about the system and that can be a tool to implement for the energy management. In particular, the methodology is applied to an educational building, the Polytechnic of Turin, focusing on a defined subsection of the system that includes energy-intensive loads and a photovoltaic production plant. The analysis follows two parallel paths, taking into account first, the load-side and then the production-side of the domain. The load-level analysis identifies typical profiles of consumption - with correspondent external conditions - of a chiller unit and an independent building, by means of an hierarchical clustering technique and a classification tree. Then, the focus is on the baseload of each profile, intended as the minimum value of demand that is always present, in order to find reference power ranges that are used to define a Key Performance Indicator, that ranks the daily energy-related behaviour. At this point, the energy waste of the loads is detected with a comparison between the actual consumption and a simulated one, considering improved values of baseload power. The production-level analysis, instead, consists in the development of an Artificial Neural Network for the forecast of the power production of the photovoltaic plant; the results of the neural network are then used to develop an anomaly detection algorithm in order to automatically found faulty operating conditions of the system, providing a daily warning that distinguishes between strong and possible anomalies. Finally, a predictive maintenance procedure is proposed with the aim to recommend extraordinary maintenance actions if a series of anomalous day are consequently reported.