Non-Intrusive Appliance Load Monitoring (NIALM): Review and Outlook *

Transcription

1 Fraunhofer Center for Sustainable Energy Systems Non-Intrusive Appliance Monitoring (NIALM): Review and Outlook * Michael Zeifman, Ph.D. Kurt Roth, Ph.D. International Conference on Consumer Electronics Las Vegas 1 January 211 * Supported in part by US DOE through a subcontract with Lawrence Berkley National Laboratory (LBNL). We are grateful to Rich Brown of LBNL for helpful discussions. Fraunhofer USA 211

2 Fraunhofer: A Leading Force in Applied R&D Fraunhofer Gesellschaft: Europe s largest contract research organization 17, employees Annual research budget: US$2.2B 59 institutes provide contract R&D and certification services for clients Research Themes: health, security, communication, mobility, energy, environment Fraunhofer USA: 7 Centers, >2 employees, Annual research budget >$5M Fraunhofer CSE is located in Cambridge, MA Three focus areas: Building Energy Efficiency PV Modules Smart Grid Fraunhofer USA 211 2

3 What Is NIALM? Non-Intrusive Appliance Monitoring A.k.a. Non-Intrusive Monitoring Power, W 4 2 Coffee Maker 7: 7:3 8: 8:3 9: 9:3 Po ower, W : 7:3 8: 8:3 9: 9:3 Time Power, W Power, W 1 5 Refrigerator 7: 7:3 8: 8:3 9: 9:3 4 2 TV 7: 7:3 8: 8:3 9: 9:3 Time Main breaker/circuit level Data acquisition (hardware) and disaggregation algorithms (software) Fraunhofer USA 211 3

4 NIALM: Interest Number of US granted patents exponential growth Patents/year Year Both large (Belkin, GE, IBM, Intel,) and small (4home, Emme, Enetics, Navetas) companies Fraunhofer USA 211 4

5 NIALM: Motivation Energy saving potential Residential buildings consume ~ 2% of primary energy in USA Electricity constitutes ~ 7% of residential energy consumption Feedback on energy usage can result in 5-15% saving Smart Grid unique business opportunity ~ 8 M smart meters in USA are projected by 213 Hardware can be used by NIALM Non-intrusive manual alternative to home automation Time-of-use electricity pricing Transition from fixed to time-dependent pricing in Europe and North America More people are concerned with energy monitoring and management Fraunhofer USA 211 5

6 How Does NIALM Work? Data Acquisition Appliance Signal Features Disaggregation Algorithm NIALM Fraunhofer USA 211 6

7 Data for NIALM: Sampling Rate Typical daily number of switching events in a household: up to ~ [Baranski, 24].1 Hz minimum sampling rate Inexpensive hardware: 1 Hz 15 More expensive hardware: > 5 khz.1.5 Power, W 1 Current, A :35 22:4 22:45 22:5 22:55 23: 23: Time, s Fraunhofer USA 211 7

8 NIALM: Types of Appliances Permanent On/off Multi State Variable Fraunhofer USA 211 8

9 I. Low Sampling Rate, Features: Power Change Real Power Reactive Power 2 2 Voltage Voltage Current Current Power 1 Power Time Time Fraunhofer USA 211 9

10 I. Power Change: Basic Method Hart, 1992 (MIT method) On/off appliances Multi State Variable Permanent Same- Appliances Accuracy: ~8% Fraunhofer USA 211 1

11 I. Change of Real Power Only Baranski and Voss, Hart s Method: single - P single + P Baranski s Method: multiple - P i multiple + P i Baranski s algorithm error term minimization (optimization by genetic algorithm) On/off appliances Multi State Variable Permanent Same- Appliances Accuracy: ~85% Fraunhofer USA

12 I. Change of Power + Additional Features Albicki and Cole, 1998 Edges and slopes as additional features On/off appliances Multi State Variable Permanent Same- Appliances Accuracy: ~N/A Fraunhofer USA

13 II. High Sampling Rate, Harmonics as Features Basic Idea: Fourier Transform 1 Current Signal waveform rrent Cur Time, s 1 1 st 3 rd Signal harmonics Time, s FFT Frequency, Hz Fourier transform Fraunhofer USA

14 II. Features: Harmonics for Transients (1/2) Laughman et al., 23 (MIT).1.8 On/off appliances.6.4 Current Multi State Time, s 1) Detect transients 2) Subtract pre-transient from posttransient 3) Take Fourier transform and compare with template 4) Set of harmonics = power envelope Variable Permanent Same- Appliances Accuracy: ~N/A Fraunhofer USA

15 II. Features: Harmonics for Transients (2/2) Lee et al., 25 (MIT) On/off appliances P harmonic Multi State Variable P total 1) Strong correlation between harmonics and power draw 2) Estimate harmonics then estimate load Permanent Same- Appliances Accuracy: ~N/A Fraunhofer USA

16 II. Features: Harmonics for Steady and Transients Srinivasan et al., 26 (Singapore) Basic idea: Set of harmonics for all combinations of appliances 1) Obtain Fourier transform continuously 2) Compare with templates On/off appliances Multi State Variable Permanent Harmonic number Same- Appliances Accuracy: 77% (phone charger) 99% (PC) Fraunhofer USA

17 II. Features: Fourier Transform of Noise (1/2) Patel et al., 27 (Ga. Tech.) Basic idea: appliance connected to a socket induces noise (electromagnetic interference) in another socket On/off appliances Multi State Variable 1) Measure signal at high rate (5 khz) 2) Calculate Fourier transform continuously 3) Compare with templates Permanent Same- Appliances Accuracy: 85-9% Fraunhofer USA

18 II. Features: Fourier Transform of Noise (2/2) Patel et al., 21 (Now U. Wash.) Basic idea: switch mode power supplies (SMPS) generate distinctive noises On/off appliances Multi State Variable Distinctive features ( applicable across homes ) Yet features variable enough to differentiate between similar devices in a home Permanent Same- Appliances Accuracy: up to 93.8% Fraunhofer USA

19 II. Features: Beyond Harmonics / Fourier Transform Liang et al., 21 (China Light & Power) Waveform: On/off appliances Multi State Power waveform Variable Permanent Same- Appliances Accuracy: 92.7% Fraunhofer USA

20 NIALM Algorithms (1/3) Single-matching (pattern recognition) Example: Waveform On/off appliances Current, a.u. Before Time, ms - Current, a.u. After Time, ms Multi State Variable = Current, a.u Time, ms Pattern recognition methods used: N-nearest neighbor Bayes classifier Neural networks Permanent Same- Appliances Fraunhofer USA 211 2

21 NIALM Algorithms (2/3) Start Multiple-matching (Optimization) Stop On/off appliances 8 Power, W Power, W Refrigerator Coffeemaker? Power, W Multi State Variable Power, W : 7:3 8: 8:3 9: 9:3 Time Permanent TV Optimization methods used: Least squares Integer programming Genetic algorithms Same- Appliances Fraunhofer USA

22 NIALM Algorithms (3/3) True disaggregation A.k.a. blind source separation On/off appliances Possible methods: Independent Component Analysis (ICA) Singular Value Decomposition Non-negative Matrix Factorization Yet has to be tried! Multi State Variable Permanent Same- Appliances Fraunhofer USA

23 Promising Directions Combination of various orthogonal features On/off appliances Combination of orthogonal algorithms Only one publication so far (Liang et al., 21) The curse of false positives Use of techniques known elsewhere Information Fusion Dempster-Shafer Evidence Theory Multi State Variable Permanent Same- Appliances Fraunhofer USA

24 NIALM Accuracy Metrics Only one paper discusses: Liang et al., 21 Can be appliance-related or overall Example: Devices were turned on 8 times over monitoring period. 7 out of 8 detected as events (no classification yet) AND also 15 false positive detections. Out of 7, 5 were correctly classified. What is the accuracy??? Detection accuracy: 5/(7+15) = 58.8% Classification accuracy: 5/7 = 71.4% Overall accuracy: 5/8 = 62.5% Fraunhofer USA

25 NIALM Accuracy Metrics ROC We suggest: Receiver operating characteristic (ROC) Fraunhofer USA

26 Conclusions and Future Trends No complete NIALM solution is available No complete set of robust, widely accepted appliance features is available Using more orthogonal features improves accuracy, albeit with higher false positive rates Using several orthogonal disaggregation algorithms may improve accuracy, but optimal fusion needs to be implemented ROC curves could be used for algorithm benchmarking rather than the ambiguous accuracy Fraunhofer USA

27 Our (Fraunhofer CSE) Research Directions Exploration of device signatures to identify new orthogonal features Development of identification/tracking algorithms suitable for the new features Optimal fusion of new algorithm(s) with the algorithms available for traditional features Testing and benchmarking device, whole home & branch circuit levels Exploration of transient event detection techniques Exploration of signal processing techniques for reconstruction of sources Results to be presented at: Viterbi Algorithm with Sparse Transitions (VAST) for Nonintrusive Monitoring, IEEE Symposium Series on Computational Intelligence, Paris, 211 Nonintrusive appliance load monitoring (NIALM) for energy control in residential buildings, Energy Efficiency in Domestic Appliances and Lighting Conference, Copenhagen, 211 Fraunhofer USA