Generic occupant behavior modeling for commercial buildings

Markovic, Romana; van Treeck, Christoph Alban (Thesis advisor); Hong, Tianzhen (Thesis advisor); Azar, Elie (Thesis advisor)

Aachen (2020)
Dissertation / PhD Thesis

Dissertation, Rheinisch-Westfälische Technische Hochschule Aachen, 2020


Human-building interactions are driven by a complex combination of social, psychological and physiological factors. As such, occupants’ energy consumption related actions can not be addressed using analytical approaches, as conventionally adopted in building performance simulation (BPS) and building automation systems (BASs). An additional degree of occupant behavior (OB) modeling complexity comes from occupants’ individuality and diversity. As a consequence, occupants’ energy consumption related behavior may be highly variant, even in case of similar settings and indoor environmental quality (IEQ). Given a large occupant population, this result in complex, yet contradictory requirements on diversity representation and model scalability. The aim of this thesis is to develop models towards more reliable OB modeling. In particular, it is aimed to gain knowledge towards more generic OB modeling, that could be applicable for a number of diverse occupants in different commercial building settings. For that purpose, the methodological focus is on machine learning (ML) methods using physical monitoring data. In order to obtain and quantify models’ generalization to alternative occupants’ and buildings, the building-wise modeling paradigm is followed through the major part of this thesis. Firstly, the potential of the time-independent OB in terms of manual window openings is explored. The modeling is conducted using conventional machine learning and deep learning classification approaches. The data imbalance is identified as a key modeling challenge. The obtained results show that the random forest based classification and developed deep learning model can reliably represent window opening behavior in given settings. As an alternative to the time-independent modeling, the sequence based modeling of OB is explored. The modeling objectives is defined to be adaptive and non-adaptive OB in commercial settings. The resulting target functions are window states modeling and miscellaneous electric loads (MELs). The sequential nature of proposed models is represented by including the time-series of past IEQ and OB measurements as the model inputs. The results show that the model formulations where the short- and long-term past of IEQ and OB data are used as inputs resulted in improved models’ performance, when compared to the alternative, established methods. Conclusively, the imbalanced properties of OB data and limited models’ applicability to alternative buildings are identified as the major limitations of the current OB modeling practices, that are addressed in the scope of this thesis. Finally, the presented models lead to more accurate yet scalable OB modeling and they show the practical potential for the inclusion in BAS and BPS as end-use real-life applications.