The main aim of the paper is to introduce a set of metrics for evaluating the event detection algorithms in NILM and to provide a dataset which enables others to test their event detection algorithms. The paper also briefly discusses the methods used for event detection in the NILM (for more information about NIALM please see Oliver's blog). According to the paper, there are three main categories which aim to detect explicit change-points in a time-series signal. The categories are listed as follows.
- Expert Heuristics: The detection of step-like changes present in the power consumption signal with the use of the standard deviation of the fixed-size set of examples.
- Probabilistic Models: As an example, the Generalized Likelihood Ratio(GLR) is introduced. It requires the parameters including the window length , the power change and a threshold so as to calculate decision statistic.
- Matched-Filters: The matching of a known signal (mask) with an unknown one in order to detect the mask in the unknown signal.
The event detection metrics introduced for event-based NILM approach are True Positive Rate, True Positive Percentage, Total Power Change, Average Power Change and Score Function. In order to show the validity of the metrics, the paper uses a modified generalized likelihood ratio(GLR). The GLR is tested with a fully-labeled dataset for BLUED dataset which is available here.
The paper concludes that the total power change metric outperforms the other three metrics which are True Positive Rate, True Positive Percentage and Average Power Change.
