This paper aims therefore at providing the first steps for a user-orientated control system for mechanical ventilation. As stated by Carbonare [3], the correct characterization of the user-dependent variables in residential buildings plays a huge role to obtain improvements on the current state-of-the-art. We define a methodology for future research and will be tested on two variables (window opening and indoor temperature), which were already studied by many researchers on the field of building simulation [4][5][6]. For reasons of practicality, only the window opening variable will be described, and in the case of the indoor temperature only the results will be presented.
BackgroundClustering is the process of classifying data into different groups, aiming at finding similarities among them. A cluster is then defined as a subset of objects in the database that belong to the same group. Similarity is often calculated through distance measures. The main challenges of a clustering process are [7]:The aim of this paper is to investigate possible patterns of the occupant behavior in residential buildings. Measurements were taken in multi-family buildings where several occupant-related variables were recorded. We chose and compared two different clustering methods: whole time series and features clustering (k-means algorithm). The mentioned methods were performed selecting two variables (window opening and indoor temperature) and tested with supervised learning methods. Results suggest that features clustering can perform better than whole time series. The representation of the occupant behavior through features is meant to be applied in future work regarding the optimization of control strategies in ventilation systems.
Eingruppierung des Nutzerverhaltens inWohngebäuden: ein Me thodenvergleich. Ziel dieser Arbeit ist es, mögliche Muster des Nutzerverhaltens in Wohngebäuden zu erkennen. Von einem Mehrfamiliengebäude wurden zunächst Daten, die auf das Nutzerverhalten schließen lassen, über einen Zeitraum von zwei Jahren erfasst. Zur Auswertung wurden zwei ClusteringMe thoden verglichen: das Clustering von Zeitreihen und die Features ClusteringMethode (mittels kMeansAlgorithmus). Die gen annten Methoden wurden hauptsächlich mit zwei Variablen (Fens teröffnung und Innentemperatur) durchgeführt, und mit Machine LearningMethoden getestet. Die Ergebnisse weisen auf eine bessere Performance der Features ClusteringMethode gegenüber dem Clustering von Zeitreihen hin. Die Darstellung des Nutzerverhaltens durch Feature Clustering wird in einer zukünftigen Arbeit als Teil einer Optimierung der Regelungsstra tegien von Lüftungsanlagen genutzt.