Bachelor's thesis Markus Schraven
Comparison of clustering algorithms for typical load profiles for the optimal design of building energy systems
The buildings sector causes 40 % of the total energy consumption in Germany. A demand controlled
generation is one opportunity to yield for great energy savings and can be achieved with an
optimally designed building energy system. As high-resolution input data and the combination of
many different components increase the complexity of design computations, long computing times
are the consequence.
A common practice in reducing the computing time is a chronological classification of the input data
into months, while representing each month’s data by its mean value. In this thesis, five clustering
methods are presented, which aggregate data based on their similarity and reduce each aggregated
group by choosing - method of k-medoid and k-center - or calculating a representative for each
group, which is done by k-means or k-means* respectively and k-median. Furthermore it is investigated
to what extend the design results of the aggregated input data produced by these clustering
methods correspond to the design results of the reference case, where the input data was not aggregated.
For the input data three different housetypes are considered, a single-familiy, a multi-family
house and an apartment building, as for the aggregation group quantities of 1 to 20 are taken into
First results show that using the investigated clusteringmethods always yields a better representation
of the reference case than taking the monthly means. Calculated representatives more often lead
to the same combination of components as in the reference case than methods, that use original
data as representatives. For instance for the multi-family dwelling k-means and k-means* find the
reference setup in 17/20 and 16/20 cases where k-medoid and k-center in merely 11/20 and 7/20.
On the other hand methods with an original representative choose more realistic dimensions of the
component which simultaneously reduces the deviation in the total cost to the reference scenario.
Beside the quality in representing the original data the size of the building energy system is crucial
when using any aggregation: The smaller the system, the more sensible it becomes for deviations
and the more deviating are design results with aggregated and not aggregated data; the higher the
aggregation’s quality the better it can detect if a general demand exists which results in the installion
of a component. In a nutshell, if the system’s robustness is high enough to tolerate the deviation on a
specific aggregation’s quality level, the aggregation constitutes a good trade-off between a decreased
quality compared to the reference case and decreased computing times in turn.