Authors:
Dejan Radovanovic
1
;
2
;
Maximilian Schirl
2
;
Andreas Unterweger
1
;
2
and
Günther Eibl
2
Affiliations:
1
Paris Lodron University of Salzburg, Salzburg, Austria
;
2
Center for Secure Energy Informatics, Salzburg University of Applied Sciences, Puch bei Hallein, Salzburg, Austria
Keyword(s):
Load Profile Analysis, Supervised Machine Learning, Evaluation Methodology, Privacy.
Abstract:
Energy consumption data from smart meters has been shown to infer socio-demographic characteristics, which impacts privacy. However, the impact of time granularity on the ability to classify such characteristics has not yet been investigated in existing literature. In this paper, we answer this question by analyzing a dataset of more than 1,000 households over one year. We obtain three main findings: (i) While a coarser time granularity leads to decreased classification performance, we find that, unexpectedly, classification performance only varies insignificantly within two relatively large granularity intervals. For example, one-hour granularity exhibits nearly the same classification performance as 15-minute granularity. This indicates that, depending on the use case, data collection can be minimized, as any resolution between 15 minutes and one hour can be used without significantly impacting prediction performance. (ii) We propose a new evaluation methodology where an interpreta
ble classification algorithm can predict a household’s socio-demographic characteristics from a load profile of a single, arbitrary week of the year. Compared to existing methodologies, where training and testing data are sampled from a single known week, using arbitrary weeks as input makes classification harder, thus requiring more sophisticated classification algorithms. (iii) We present such an interpretable classification algorithm, which outperforms those that train and evaluate classifiers separately for each week. At the same time, our algorithm exhibits a comparable performance to approaches that require a load profile of the whole year instead of a single, arbitrary week.
(More)