Bachelor's Thesis Pascal Stammer
Study of unsupervised machine learning algorithms for topology detection on complex building energy systemsCopyright: EBC
The building sector contributes a major share to the total energy consumption, which is rising year by year. However, many buildings are already equipped with measuring and control technology that can be used for automation systems, which are a promising source for energy saving. Topology detection algorithms can improve the understanding and control of complex energy systems. Clustering multivariate time series of energy systems for topology detection is a rather unexplored field of research. Therefore, the goal of this thesis is to advance the research in this field, by studying the performance of two model based, unsupervised clustering machine learning algorithms when applied to several real BACS datasets. In addition, an insight on the evaluation of the results of algorithms, especially when examining real datasets, is given. And consequently, two different evaluation approaches are presented and analyzed.
This thesis is based on the Takeshi project, in which the Takeshi algorithm was developed and tested. The studied algorithms, TICC and MASA were able to outperform the Takeshi algorithm by over 20% in the F1 score, both reaching values of around F1 = 0.6, depending on the used dataset. Both algorithms were applied on two different datasets of the E.ON ERC Main Building as well as a dataset from another real energy system, named Werk 3. While the results of the TICC and MASA algorithm are promising, there is still improvement to be made in the utilization of these algorithms, to reach their full potential. Therefore, different approaches for optimizing the usage of the studied algorithms in future applications, are proposed.