PaperlayerView exact revision on arXiv

Source-linked AI summary

Non-Intrusive Signature Extraction for Major Residential Loads

M. Dong, P. C. M. Meira, W. Xu, C. Y. Chung

arXiv:1804.11049v1eess.SPcs.AIcs.CE

TL;DR

Smart-meter data could support appliance-level energy tracking, but non-intrusive extraction of home-specific load signatures remains an identified research problem. The paper proposes a multistep event-based method and reports reconstructed cycles and signatures from real-house data, including extraction of appliances with incompletely labeled activity.

Abstract

from arXiv · show

The data collected by smart meters contain a lot of useful information. One potential use of the data is to track the energy consumptions and operating statuses of major home appliances.The results will enable homeowners to make sound decisions on how to save energy and how to participate in demand response programs. This paper presents a new method to breakdown the total power demand measured by a smart meter to those used by individual appliances. A unique feature of the proposed method is that it utilizes diverse signatures associated with the entire operating window of an appliance for identification. As a result, appliances with complicated middle process can be tracked. A novel appliance registration device and scheme is also proposed to automate the creation of appliance signature database and to eliminate the need of massive training before identification. The software and system have been developed and deployed to real houses in order to verify the proposed method.

I. INTRODUCTION

Smart-meter data motivate non-intrusive load monitoring, but extracting home-specific appliance signatures is a distinct, previously uninvestigated problem. The paper proposes a five-step method that derives signatures from detected power-change events.

I. INTRODUCTION

  • Home-specific signatures may improve identification accuracy over generic signatures and can support load-condition monitoring.The paper presents non-intrusive collection as preferable to intrusive signature collection for ordinary homeowners.Source: §I · p. 1
  • Non-intrusive signature extraction is framed as establishing appliance-specific signatures from multiple days of power data, a distinct problem not previously investigated as such.The introduction contrasts this task with non-intrusive identification of load activities.Source: §I · p. 1
  • The proposed method detects power-change events and proceeds through event detection, filtration, clustering, association, and signature extraction.The method is validated using real-life data and a laboratory experiment.Source: §I · p. 1

II. EVENT FILTRATION

Event filtration narrows the search for appliance events using electrical characteristics and operating-time context. Its filters must retain enough candidate events for later processing, while their ranges may need local adjustment.

II. EVENT FILTRATION

  • Appliance events are captured through edge detection of discontinuities in power; cycles may contain an ON/OFF pair or multiple intermediate state changes.The method uses these event characteristics to roughly locate appliance activity.Source: §II · p. 2
  • Filtration restricts candidate events by active and reactive power, harmonic content, spikes, phase, and search time.The method verifies ON events as potential suspects, since many transient features such as spikes accompany ON events.Source: §II · pp. 2–3
  • Electrical ranges reflect appliance attributes: a resistive kettle is described as high-power with low reactive power and nearly zero harmonics, while motors and switch-mode electronics have different signatures.Different stove heating elements may also be treated as different appliance types because their signatures differ.Source: §II · pp. 2–3
  • Searching less-overlapped operating periods can reduce interference, and time windows from multiple days are joined into a larger data piece where target events are expected to be denser.Examples include searching for a fridge from 2:00 AM–5:00 AM and a microwave before meals.Source: §II · p. 2, Figure 2 · p. 2
  • The ON-event ranges are generic to the Edmonton research region, expanded at least 20% beyond measurements, and may require local updating across regions.For example, the measured microwave power range of 1200–2000 W was expanded to 800–2500 W in the table.Source: §II · pp. 2–3
  • Filters should remain permissive enough to include sufficient appliance events, even if that allows some unrelated events into the suspect set.Insufficient event samples can make later procedures fail or produce inaccurate results.Source: §II · p. 3

III. EVENT CLUSTERING

Clustering separates authentic appliance events from unrelated or corrupted events that pass filtration. The method assumes target events dominate the suspect set and selects the largest cluster as authentic.

III. EVENT CLUSTERING

  • The clustering step distinguishes authentic events from unrelated events that can enter a suspect group because filtration uses loose ranges across multiple days.Authentic events are the real events belonging to the target appliance.Source: §III · p. 3
  • The method assumes target-appliance events greatly outnumber other events in a suspect group, and it is not intended to handle abnormal cases where this fails.Simultaneous appliance events can also create corrupted events, though these are described as normally much less frequent than authentic events.Source: §III · p. 3
  • The largest event cluster is treated as authentic, while smaller clusters are discarded as likely other appliances or noise.Filtration first reduces the event space, allowing clustering to be faster and more accurate than clustering all daily events.Source: §III · p. 3

A. Mean-Shift Clustering

Mean-shift clustering is used because the number of event clusters is not known in advance. It clusters normalized P/Q/THD features with a tuned bandwidth.

A. Mean-Shift Clustering

  • Mean-shift is selected as a robust non-parametric method because the number of appliance clusters is unknown beforehand.Methods such as K-means are therefore described as inapplicable in this setting.Source: A. Mean-Shift Clustering · p. 4
  • The clustering features are P/Q/THD, normalized to the range 1 to 100, with mean linkage used to determine distance between clusters.Source: A. Mean-Shift Clustering · p. 4
  • The mean-shift bandwidth is set between 5 and 20 to account for estimated signature variation from system-voltage fluctuations.Source: A. Mean-Shift Clustering · p. 4

B. Weight-Based Clustering

Weight-based clustering assigns appliance-category-specific importance to electrical features, addressing mean-shift’s equal weighting of attributes. In the comparison, it excludes two fan events that mean-shift incorrectly groups with the fridge.

B. Weight-Based Clustering

  • Appliance-category-specific feature weights emphasize informative attributes, such as active/reactive power for fridges and negligible reactive power for stoves.This weighting addresses mean-shift’s treatment of all attributes as equally important in Euclidean space.Source: B. Weight-Based Clustering · p. 4
  • The clustering iteratively merges events similar to cluster means, then merges similar cluster means until no cluster changes.Similarity uses weighted electrical attributes and can be judged against a defined threshold, such as 0.8.Source: B. Weight-Based Clustering · p. 4
  • When within-group power variance changes, the method adjusts bandwidth or similarity threshold to favor cluster separation or fusion.Higher variance calls for lower mean-shift bandwidth and higher weight-based threshold; lower variance reverses those adjustments.Source: B. Weight-Based Clustering · p. 4
  • Weight-based clustering excludes two fan events from the fridge group, whereas mean-shift mistakenly includes them.The paper attributes the difference to down-weighting harmonic content for a linear/reactive fridge, since similar THD values confound the methods.Source: B. Weight-Based Clustering · p. 4

IV. EVENT ASSOCIATION

Event association identifies which non-ON events belong to an appliance so its full representative operating cycles can be reconstructed. It classifies event relationships by their strength and occurrence across segments beginning with authentic ON events.

IV. EVENT ASSOCIATION

  • Association recovers events beyond the ON event, including OFF and intermediate events, to reconstruct full appliance cycles and patterns.This is especially important for multi-stage appliances such as furnaces, whose ON event is only a small part of the operating cycle.Source: §IV · pp. 4–5
  • Events are classified as single, repetitive, occasional, or unrelated according to how strongly and how often they accompany an appliance’s operation.Single and repetitive events are strongly associated; occasional events may be absent, while unrelated events come from another appliance.Source: §IV · p. 5
  • Association is assessed across segments that begin with authentic ON events, using event counts and occurrence frequency to distinguish event types.Segments are sized to include associated events without admitting too many unrelated events; a reasonable length is 1.5 times the appliance’s average operating duration.Source: §IV · p. 5
  • Count-based criteria are corrected using confidence and additional conditions to accommodate mistaken clustering or events missed by segmentation.The passage gives typical values of b = 0.3 and c = 0.8.Source: §IV · p. 5, Table IV · p. 5

V. VERIFICATIONS AND DISCUSSIONS

In house #1, reconstructed cycles were compared with manually labeled reference cycles and found similar, with acceptable electrical-signature accuracy. The evaluation also found that authentic events dominated suspect-event clusters.

A. Verification and Discussions Based on Real House #1’s Data

B. Verification and Discussions Based on Real House #2’s Data

The method was evaluated in a second residential house, where it found eight major appliances and produced electrical signatures described as quite accurate against labeled reference cycles.

B. Verification and Discussions Based on Real House #2’s Data

C. Verification and Discussions Based on Real House #3’s Data

On REDD house #3 data, the method extracted five major appliances, including washer cycles that were unlabeled in the dataset. Reconstructed cycles had very good electric accuracy under Edmonton-derived filtration conditions, which remained viable for the Boston-area house.

C. Verification and Discussions Based on Real House #3’s Data

  • The house #3 evaluation used 1,427,284 seconds of REDD data, with many middle hours missing and some appliance labels unclear.The dataset's sampling rate was 275 points per cycle.Source: C. Verification and Discussions Based… · p. 7
  • Table VIII reports very good electric accuracy for reconstructed cycles, and Edmonton-based filtration conditions remained viable for the Boston-area house.The authors suggest appliances may share signature ranges in similar geographic regions, such as northern North America.Source: Table VIII · p. 8, C. Verification and Discussions Based… · p. 8
  • Five major appliances were extracted from REDD house #3 despite missing activity periods and unclear appliance labels, and the method located and relabeled unlabeled washer cycles.The dataset covered roughly 16 days; stove, oven, and kettle activities were among those not recorded in some periods.Source: C. Verification and Discussions Based… · pp. 7–8

D. Verification of Event Association Based on Laboratory Data

Laboratory tests assessed whether event association could distinguish a space heater’s own changing operation from events caused by other appliances. Across four scenarios and 12 data segments, the resulting association pattern matched the heater’s physical behavior; the overall approach requires more than a week of input data for accurate signatures.

D. Verification of Event Association Based on Laboratory Data

  • The heater experiment placed a lamp and microwave on the same power supply bar to introduce interference into the aggregated signal.The four scenarios varied whether the heater’s swaying function was active and whether the other appliances switched on during operation.Source: D. Verification of Event Association… · p. 8
  • Clustering covered four heater scenarios across 12 data segments, and Table IX reports the event-association judgment.The scenarios combined heater operation with optional swaying and events from a microwave and lamp.Source: D. Verification of Event Association… · p. 8, Table IX
  • The determined event-association pattern was completely consistent with the heater’s physical observations across the tested scenarios.The tests included repetitive heater events, an occasional sway function, and interfering microwave and lamp events.Source: D. Verification of Event Association… · pp. 8–9
  • Accurate electric and event-pattern signatures require sufficient input data, specified as more than a week, after which existing NILM methods can be used without intrusive training or registration measurements.The authors describe combining NISE with an NILM method as a route to non-intrusive load monitoring.Source: D. Verification of Event Association… · p. 9

VI. CONCLUSION

The paper frames non-intrusive load-signature extraction as a distinct NILM problem and proposes unsupervised extraction from meter-side data with almost no user effort. It evaluates extracted signatures using three residential houses and a laboratory experiment, while identifying smaller or unique appliances and geographically updated filtration conditions as future work.

VI. CONCLUSION

  • The proposed unsupervised approach uses meter-side data to automatically extract appliance signatures with almost zero user effort.It uses event filtration, clustering, and association to identify authentic events and reconstruct appliance operation cycles containing electric and event-pattern signatures.Source: §VI · p. 9
  • The paper addresses non-intrusive signature extraction as a distinct problem from the load-activity identification or classification emphasized by prior NILM studies.The authors state that signature extraction had remained unsolved and that intrusive extraction can significantly affect NILM application in ordinary households.Source: §VI · p. 9
  • The approach was verified offline using data from three residential houses and a laboratory experiment, with extracted-signature accuracy described as generally satisfactory.Both electric and event-pattern signatures were tested, analyzed, and discussed.Source: §VI · p. 9
  • The study focuses on major appliances; smaller or unique appliances are left for future study.The authors also propose updating event-filtration conditions using extensive measurements from different geographic regions.Source: §VI · p. 9
Loading 1804.11049v1…