Data-Driven Methods for the Management and Assessment of Final Energy Uses for Indoor Comfort

This thesis investigates data-driven methods for the management and assessment of final energy uses associated with indoor comfort, with particular attention to visual comfort in daylight-linked lighting systems and thermal comfort in radiant heating and cooling environments. The research examines how advanced data-driven models can support the estimation of comfort-related variables, the evaluation of sensor positions, and the interpretation of system behaviour in both measured and simulation-based building applications. A unified methodological framework was adopted throughout the work, combining data curation, model development, comparative evaluation and sensing analysis across a set of interconnected case studies carried out mainly in the SolarLab and in classrooms T120 and T220 of Building 9 at the University of Palermo, Italy.The first research stream addresses workplane illuminance estimation and ceiling photosensor placement for daylight-linked lighting control. In the SolarLab measured case study, the results show that workplane illuminance can be estimated accurately from ceiling photosensor measurements and a limited set of daylight-related variables. Among the tested models, boosting-based methods provided the most effective balance between predictive performance and practical applicability, while convolutional and recurrent deep learning architectures also achieved competitive results. The analysis further showed that model comparison should not rely only on full-range regression accuracy, since behaviour near the maintained illuminance threshold has direct implications for dimming reliability. In this environment, one ceiling photosensor position was found to be more representative than the other for workplane illuminance estimation. The same framework was extended to annual simulation datasets for classrooms T120 and T220, where preferred ceiling photosensor positions were identified for both rooms, and a complementary measured data validation confirmed the ability of the approach to distinguish among candidate sensor locations under real monitored conditions. A further contribution concerns the assessment of the actual performance of the installed daylight-linked lighting system in the SolarLab. Using a diagnostic framework based on multiple energy performance indices, the study showed that the installed system achieved reasonably good over-illumination avoidance but insufficient performance in maintaining the target workplane illuminance. This result highlights the importance of distinguishing between genuinely effective daylight exploitation and apparently low energy consumption caused by inadequate maintained illuminance.The second research stream focuses on thermal comfort modeling in a zonally controlled radiant environment. The simulation-based analysis in the SolarLab showed that accurate estimation of predicted mean vote can be achieved using a reduced but informative set of thermal variables. Within the best input configuration, the feedforward artificial neural network emerged as the strongest-performing model. The results also showed that the MRT2 sensing position provided more reliable support for thermal comfort estimation than MRT1. Overall, the thesis demonstrates that data-driven methods can support not only the estimation of indoor comfort variables, but also sensing decisions, commissioning-related evaluations, and comfort-oriented operational assessment in real building environments.