Time series clustering by extracted features
Felix Hildén · LUTPub (LUT University) · 2019
Clustering electricity consumers can give insight into how a particular consumption profile is related to other attributes of the customer, for example the residence size or whether a sauna is in use or not. Furthermore, if the profiles are separated well enough, more accurate models predicting consumption could be built. To cluster consumers, a set of static features were extracted from raw time series representing hourly energy consumption. These features were then clustered with a self-organising map. To visualise the results, background data detailing additional information on each consumer was used as categorical labels for each series. The results were promising. A set of background data produced some separation on the map, which indicates that there are tangible differences to the hourly consumption data alone.