Machine learning-based analysis of residential electricity consumption behavior for consumers and prosumers

With the shift towards a more sustainable energy system, the need for a better understanding of the behavior development over time of consumers and prosumers arises. Despite the growing penetration of smart meter infrastructure, the availability of energy usage behavior data is still limited, due to privacy and security concerns. Thus, connecting and comparing existing datasets is the key to observe the user behavior shifts as well as enhancing the utility of the available data. In the present work, a novel work!ow for combined analysis on multiple smart meter datasets is proposed, which links datasets with diferent scopes, temporal origins and speci#cations. In general, there are 4 steps: data preprocessing, clustering, location dependency check and dataset linking. First, the meteorological seasons combined with weekdays and weekends are picked for data segmentation in the data preprocessing, followed by missing value validation and normalization based on the maximum and minimum consumption value of each household. Thereafter, K-means clustering algorithm is applied to group the user behaviors, which stands out by 0.8186 Silhouette coe$cient (SIL) and 0.2884 Davies-Bouldin Index (DBI) among Fuzzy C-Means and hierarchical clustering approach. Subsequently, two validation approaches on the location dependency, cluster center correlation (0.8048) and location share among clusters (4.99 % variability), prove the minor impact of the household location on the electricity consumption behavior within Germany. Based on the location dependency check, ultimately, the combined analysis of the two datasets reveals the temporal development of the residential consumption behaviors. It shows that new technologies, especially Photovoltaics (PV), Electric Vehicles (EV) and heat pumps, have in!uence on the user behavior shift and the energy consumption level.