1. LABEL CORRECTION AND EVENT DETECTION FOR ELECTRICITY DISAGGREGATION
Mark Valovage1 and Maria Gini2
1Ph.D. Candidate, University of Minnesota
2Professor of Computer Science, University of Minnesota

2. OVERVIEW Background Contributions Conclusions
Electricity Disaggregation
Contributions
Training Sample Label Correction (Supervised Learning)
Parameter-free Event Detection (Unsupervised Learning)
Conclusions

3. ELECTRICITY DISAGGREGATION
Annual electrical waste (United States)
$70 billion/year.
1.1 gigatons pollution.
Aggregate data
“You used X KWh.”
“Pay us $Y.”
Disaggregated data
Can reduce power consumption by 15%.

4. LABEL CORRECTION Background Contributions Conclusions
Electricity Disaggregation
Contributions
Training Sample Label Correction (Supervised Learning)
Parameter-free Event Detection (Unsupervised Learning)
Conclusions

5. ELECTRICITY DISAGGREGATION LABELING ERRORS
Limiting Assumptions
Correctly labeled samples
Isolated appliances
Labels error sources
Offsets
Raw user errors
Truncated shutdowns
Contaminations/Not isolated

6. LABEL ERRORS IMPACT ON CLASSIFICATION
Back Porch Lights Samples Mean Power Increase (W)
Uncorrected
Corrected
Sample 1
229.7
331.8
Sample 2
145.04
333.15
Sample 3
144
330.74
Sample 4
143.11
N/A
Mean
165.46
331.90
Std. Dev.
37.09
0.99
Simple Mean Classification (+/- 10%)
True Positives
0/4
4/4
False Positives 3 8
False Positive Appliances (Power Phase 2)
Dining Room Lights
344.84
Master Bathroom Lights
314.52
Bedroom 1 Lights
290.19
Bedroom 1 LCD TV
152.04
Simple Mean Classification (+/-2 std)
True Positives
0/4
4/4
False Positives 8

7. BAYESIAN CHANGE DETECTION
Detects abrupt changes in 1-dimensional data
Probability based
Tracks active sequence (highest probability)
Identifies changepoint when active run sequence changes
Advantages:
Robust to noise
Does not require parameter tuning
BCD is a tool that detects abrupt changes in 1-dimensional data.

8. LABEL CORRECTION ALGORITHMS (1/2)
Naïve Proximal (NaïveProx)
Run Bayesian Change Detection on training sample
Use closest changepoints to user on/off labels
Does not account for domain-level properties

9. LABEL CORRECTION ALGORITHMS (2/2)
Power-constrained Proximal (PCProx)
Can be online or offline (OnlinePCProx/OfflinePCProx)
Run Bayesian Change Detection on training sample
Use closest changepoints to user on/off labels
Ensure new on/off labels satisfy real-power constraints

10. LABEL CORRECTION DATASET
Kaggle Belkin dataset
4 houses
Real power sampled at 5 Hz
147 appliances
462 training samples
Label errors have not been removed

11. LABEL CORRECTION RESULTS

12. EVENT DETECTION Background Contributions Conclusions
Electricity Disaggregation
Contributions
Training Sample Label Correction (Supervised Learning)
Parameter-free Event Detection (Unsupervised Learning)
Conclusions

13. UNSUPERVISED LEARNING EVENT DETECTION
No training samples
Appliances not isolated
Limitations of existing event detection methods
Require pre-processing to reduce noise
Parameters must be tuned

14. EVENT DETECTION COMPETING ALGORITHMS
Modified Generalized Likelihood Ratio (mGLR)
Change Detection
Requires parameter tuning
Dirichlet Process Gaussian Mixture Models (DPGMM)
Identifies steady-state power levels through iterative clustering
Marks events as changes between steady-states
Reduces noise through median window filtering

15. EVENT DETECTION DATASETS
BLUED (Building-Level fUlly-labeled dataset for Electricity Disaggregation)
1 house
Power sampled at 60 Hz for 1 week
REDD (Reference Electricity Disaggregation Dataset)
6 houses
Power sampled at ~0.25 Hz

16. EVENT DETECTION PERFORMANCE METRICS
Unweighted F-measure
Power-weighted F-measure

17. EVENT DETECTION RESULTS (1/2)
UNWEIGHTED POWER-WEIGHTED

18. EVENT DETECTION RESULTS (2/2)
UNWEIGHTED POWER-WEIGHTED

19. CONCLUSIONS Introduced application of Bayesian change detection
Label Correction (Supervised Learning)
Event Detection (Unsupervised Learning)
Advantages
Improves accuracy of samples for supervised learning
Eliminate the need for parameter tuning
Perform competitively against existing methods
Can be run in real time on inexpensive hardware

20. FUTURE WORK Label Correction Event Detection
Improve correction accuracy
Recover information from contaminated samples
Event Detection
Automatically compute baseline noise threshold,
Improve appliance reconstruction algorithms

21. QUESTIONS
LABEL CORRECTION AND EVENT DETECTION FOR ELECTRICITY DISAGGREGATION
Additional materials available at
Mark Valovage
Ph.D. Candidate, University of Minnesota
Research Intern, SIFT
Maria Gini
Professor, Computer Science
University of Minnesota
That concludes my presentation. I will now take your questions.