Unsupervised Disaggregation of Low Granularity Resource Consumption Time Series
The 22nd Pacific-Asia Conference on Knowledge Discovery and Data Mining (PAKDD 2018)
2018
Conference/Workshop
- Contact persons: Spiros Athanasiou , Giorgos Giannopoulos
- Relevant research project: DAIAD
Abstract.
Resource consumption is typically monitored at a single point that aggregates all activities of the household in one time series. A key task in resource demand management is disaggregation; an operation that decomposes such a composite time series in the consumption parts that construct it, thus, extracting detailed information about how and when resources were consumed. Current state-of-the-art disaggregation methods have two drawbacks: (a) they mostly work for frequently sampled time series and (b) they require supervision (that comes in terms of labelled data). In practice, though, sampling is not frequent and labelled data are often not available. With this problem in mind, in this paper, we present a method designed for unsupervised disaggregation of consumption time series of low granularity. Our method utilizes a stochastic model of resource consumption along with empirical findings on consumption types (e.g., average volume) to perform disaggregation. Experiments with real world resource consumption data demonstrate up to 85% Recall in identifying different consumption types.