Open AccessTakeshi: Application of unsupervised machine learning techniques for topology detection in building energy systems
Author(s) -
Florian Stinner,
Lukas Raßpe-Lange,
Marc Baranski,
Dirk Müller
Publication year - 2019
Publication title -
journal of physics. conference series
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.21
H-Index - 85
eISSN - 1742-6596
pISSN - 1742-6588
DOI - 10.1088/1742-6596/1343/1/012041
Subject(s) - preprocessor , computer science , energy (signal processing) , identification (biology) , data mining , set (abstract data type) , machine learning , data pre processing , range (aeronautics) , artificial intelligence , data set , algorithm , engineering , mathematics , statistics , botany , biology , programming language , aerospace engineering
Buildings rarely achieve their energy targets defined in the planning phase. Due to a lack of time and money, operators frequently fail to identify and implement energy efficiency measures in control. Knowledge about the topology of the building energy system (BES) is important for the automatic identification of energy efficiency measures, but usually does not exist in a machine-readable form. In this paper, we present an approach to detect the topology of a BES using an unsupervised learning algorithm, called Takeshi . We apply the algorithm to real and simulated time series data of a multifunctional building. This algorithm relies on a data mining approach, in which four steps are conducted, preprocessing, partitioning, sequencing and rule detection. The results obtained using the real-life data were only partly satisfactory. The best F1 score was 43,3 %, whereby the used seasons of the year had a high influence. In order to demonstrate a broader range of applications, we applied the algorithm to the simulation data. In that case, the algorithm shows significantly better results and the F1 score reached 79,6 %. We evaluate reasons for the poor performance in the case of the real BACS data set and derive possible improvements to the methodology.