Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

An Introduction to Load Disaggregation

39 views

Published on

Smart Microgrid Meetup, Mindspace Siemens AI Lab

Published in: Engineering
  • Be the first to comment

  • Be the first to like this

An Introduction to Load Disaggregation

  1. 1. An Introduction to Load Disaggregation Christoph Klemenjak
  2. 2. All roads lead to Rome ● Load Disaggregation ● Non-Intrusive Load Monitoring (NILM) ● Appliance Load Monitoring (ALM) ● Non-Intrusive Appliance Load Monitoring (NIALM)
  3. 3. Background ● Technical High School in Telecommunications and Computer Engineering (HTL) ● Bachelor in Information Technology ● Master’s in Information and Communications Engineering ○ University of Klagenfurt, Austria ○ Focus on Smart Grids and Energy Informatics ○ Thesis on Smart Meter Design and Load Classification ● First contributions in relevant conferences ● At present: PhD student in Energy Informatics
  4. 4. Agenda ● Motivation ● What is NILM? ● Appliance Modelling ● General Framework of NILM ● Recent Implementations ● Accuracy Metrics ● Open Challenges
  5. 5. Motivation ● Detailed energy insights and feedback ● Occupancy detection ● Appliance usage modelling & forecasting ● Detection of energy-leaking appliances ● Predict maintenance windows ● Cost reduction ● Optimisation of workflows D. Chen, S. Barker, A. Subbaswamy, D. Irwin, and P. Shenoy. Non-intrusive occupancy monitoring using smart meters. in Proc. of the 5th ACM Workshop on Embedded Systems For Energy-Efficient Buildings (BuildSys’13), 2013. Armel, K. Carrie, et al. "Is disaggregation the holy grail of energy efficiency? The case of electricity." Energy Policy 52 (2013): 213-234.
  6. 6. Egarter, Dominik, and Wilfried Elmenreich. "Load disaggregation with metaheuristic optimization." Energieinformatik, Karlsruhe, Germany (2015).
  7. 7. "NILM is a set of techniques used to obtain estimates of the electrical power consumption of individual appliances from measurements of voltage and/or current taken at a limited number of locations in the power distribution system of a building." Batra, Nipun, et al. “A comparison of non-intrusive load monitoring methods for commercial and residential buildings.” arXiv preprint arXiv:1408.6595 (2014).
  8. 8. Problem specification Egarter, Dominik, and Wilfried Elmenreich. "Load disaggregation with metaheuristic optimization." Energieinformatik, Karlsruhe, Germany (2015).
  9. 9. How can we model an electrical appliance? 1. Single-state appliances: Toaster, light bulb, etc. 2. Multi-state appliances: Washing machine, dishwasher, stove burner, etc. 3. Infinite-state appliances (continuously variable): Power drill or dimmer lights 4. Permanently-on appliances Christoph Klemenjak and Wilfried Elmenreich. On the Applicability of Correlation Filters for Appliance Detection in Smart Meter Readings. In Proceedings of the 2017 IEEE International Conference on Smart Grid Communications (SmartGridComm), Dresden, Germany, 2017.
  10. 10. Non-Intrusive Load Monitoring Data Acquisition Feature Extraction Inference and Learning Zoha, Ahmed, et al. “Non-intrusive load monitoring approaches for disaggregated energy sensing: A survey.” Sensors 12.12 (2012): 16838-16866. ● Data Acquisition: perform aggregate load measurement at an adequate rate ● Feature Extraction: process the raw data and detect events ■ Non-Traditional Features ■ Steady-State Features ■ Transient State Features ● Inference and Learning is the mathematical model that disaggregates the total power into appliance level signals ■ Supervised Learning ■ Semi-Supervised and Unsupervised Learning
  11. 11. Data Acquisition ● Crucial first stage ● Performance limiting characteristics: ○ Measured physical quantities (P, Q, etc.) ○ Measurement resolution (Quantisation) ○ Sampling frequency ● Potential sources of data: ○ Smart Meter → Non-intrusive measurements ○ Ambient Sensors ○ Human-machine interaction ○ Smart Plugs → Intrusive measurements Klemenjak, Christoph, Dominik Egarter, and Wilfried Elmenreich. "YoMo: the Arduino-based smart metering board." Computer Science-Research and Development 31.1-2 (2016): 97-103. YoMo Meter CV EM24
  12. 12. Feature Extraction ● Steady-state features ○ Power Change ○ Time and Frequency Domain V-I features ○ V-I Trajectory ○ Voltage Noise ● Transient-state features ○ Transient power ○ Start-Up Current Waveforms ○ Voltage Noise ● Non-Traditional features ○ Ambient characteristics ○ Correlations ○ Time of the day
  13. 13. Inference and Learning
  14. 14. Supervised Learning in NILM ● Requires labelled data sets for training ● Optimisation methods: ○ Integer programming (Bhotto 2016) ○ Genetic algorithms (Baranski 2004) ● Pattern matching approaches: ○ Artificial Neural Networks (Srinivasan 2006) ○ Support Vector Machine (Kato 2009) ○ Bayesian approaches (Srinivasarenga 2013) Klemenjak, Christoph, and Peter Goldsborough. "Non-Intrusive Load Monitoring: A Review and Outlook." arXiv preprint arXiv:1610.01191 (2016).
  15. 15. Semi-Supervised and Unsupervised Learning ● Appliance models are captured only using the aggregated load ● Three groups, as suggested by [2]: ○ Unsupervised algorithms that require unlabelled training data to build appliance model ○ Unsupervised algorithms with labelled training data ■ Learn from seen house, apply algorithm on unseen house ○ Untrained unsupervised algorithms such as [1] [1] Zhao, Bochao, Lina Stankovic, and Vladimir Stankovic. "On a training-less solution for non-intrusive appliance load monitoring using graph signal processing." IEEE Access 4 (2016): 1784-1799. [2] Faustine, Anthony, et al. "A Survey on Non-Intrusive Load Monitoring Methodies and Techniques for Energy Disaggregation Problem." arXiv preprint arXiv:1703.00785 (2017).
  16. 16. NILM based on LSTM RNN ● Recurrent Neural Network ● LSTM with power signal as input and on/off state of each appliance as output ● Novel signature method ● Claim lower time complexity compared to FHMM ● LSTM can learn long-term patterns Jihyun Kim, Thi-Thu-Huong Le, and Howon Kim, “Nonintrusive Load Monitoring Based on Advanced Deep Learning and Novel Signature,” Computational Intelligence and Neuroscience, vol. 2017, Article ID 4216281, 22 pages, 2017. doi:10.1155/2017/4216281
  17. 17. [1] Fiol Arguimbau, Albert. Algorithms for energy disaggregation. MS thesis. Universitat Politècnica de Catalunya, 2016. [2] Li, Yeqing, et al. "Energy disaggregation via hierarchical factorial hmm." Proceedings of the 2nd International Workshop on Non-Intrusive Load Monitoring (NILM). Google Scholar. 2014. House FHMM [1] Additive FHMM [2] HieFHMM [2] LSTM F1 score 1 0.45 0.749 0.854 0.950 3 0.59 0.619 0.834 0.835 4 0.43 0.417 0.424 0.856 5 0.5 0.795 0.796 0.911 6 0.44 0.391 0.820 0.925 Performance comparison on REDD
  18. 18. Denoising Autoencoders for NILM ● NILM is treated as a denoising problem ○ Clean signal is the disaggregated consumption pattern ○ Noise signal at the input ● Convolutional and fully connected layers Kelly, Jack, and William Knottenbelt. "Neural NILM: Deep neural networks applied to energy disaggregation." Proceedings of the 2nd ACM International Conference on Embedded Systems for Energy-Efficient Built Environments. ACM, 2015.
  19. 19. Bonfigli, Roberto, et al. "Denoising autoencoders for Non-Intrusive Load Monitoring: Improvements and comparative evaluation." Energy and Buildings 158 (2018): 1461-1474. Scenario Algorithm F1 in % NEP MCC seen AFAMAP + RoW 24.9 12.32 0.195 Kelly - NeuralNILM 60.4 0.829 0.62 dAE 73.6 0.76 0.661 unseen AFAMAP + RoW 17.3 6.23 0.089 Kelly - NeuralNILM 50.8 1.27 0.608 dAE 67.8 0.81 0.683 Denoising Autoencoders vs. AFAMAP
  20. 20. Comparison of NILM algorithms ● PALDi by Egarter et al. (2015) ● FHMM with context-based features by Paradiso et al. (2016) ● Sparse Viterbi by Makonin et al. (2015) ● Denoising Autoencoder by Kelly et al. (2015) ● DCNN for regression by Kelly et al. (2015) ● RNN LSTM with conv. input layer by Kelly et al. (2015) ● RNN LSTM by Mauch et al. (2015) ● Seq2point CNN by Zhang et al. (2017)
  21. 21. Feature selection Accuracy No training Real-time capabilities Scalability Generalisation PALDi ✓ ✓ × ✓ × - FHMM with context-based features ✓ - ✓ ✓ ✓ - Sparse Viterbi ✓ ✓ × - × - Denoising Autoencoder ✓ ✓ × ✓ ✓ ✓ DCNN for regression ✓ ✓ × ✓ ✓ ✓ RNN LSTM with conv. input layer ✓ × × ✓ ✓ × RNN LSTM ✓ - × ✓ ✓ ✓ Seq2point CNN ✓ ✓ × - ✓ ✓ Taken from: Nalmpantis, Christoforos, and Dimitris Vrakas. "Machine learning approaches for non-intrusive load monitoring: from qualitative to quantitative comparation." Artificial Intelligence Review (2018): 1-27..
  22. 22. (Some) Accuracy Metrics Comparison between observed aggregate and reconstructed signal: ● Mean-squared error - MSE ● Root-mean-squared error - RMSE ● Normalised error in assigned power - NEP ● ... Typical classification metrics: ● Accuracy ● F1 - Score ● Total energy correctly assigned - TECA ● ….
  23. 23. Datasource Complexity Accuracy TECA F1-score N PALDi REDD House 1 M = 23.34 m = 4.39 m = 0.73 std = 0.24 - - 6 FHMM with context-based features Tracebase M = 10.83 m = 6.17 - - - 6 Sparse Viterbi REDD House 1 M = 6.09 m = 2.83 - m = 0.95 - 5 Denoising Autoencoder UK-DALE seen house M = 10.83 m = 6.17 m = 0.9 std = 0.08 - m = 0.55 std = 0.18 6 DCNN for regression UK-DALE seen house M = 10.10 m = 4.75 m = 0.94 std = 0.07 - m = 0.58 std = 0.15 6 RNN LSTM with conv. input layer UK-DALE seen house M = 10.10 m = 4.75 m = 0.68 std = 0.29 - m = 0.39 std = 0.28 6 RNN LSTM REDD house 1 M = 10.10 m = 4.75 - - m = 0.78 std = 0.12 3 Taken from: Nalmpantis, Christoforos, and Dimitris Vrakas. "Machine learning approaches for non-intrusive load monitoring: from qualitative to quantitative comparation." Artificial Intelligence Review (2018): 1-27..
  24. 24. Open Challenges ● Generalisation of NILM algorithms ● How good is good enough? ● Reproducibility of experiments and results ● Common accuracy metrics and data sets ● Definition of complexity ● Available toolkits need to be updated ○ NILMTK ○ NILM-eval
  25. 25. Literature 1. Nalmpantis, Christoforos, and Dimitris Vrakas. "Machine learning approaches for non-intrusive load monitoring: from qualitative to quantitative comparation." Artificial Intelligence Review (2018): 1-27. 2. Bonfigli, Roberto, et al. "Denoising autoencoders for Non-Intrusive Load Monitoring: Improvements and comparative evaluation." Energy and Buildings 158 (2018): 1461-1474. 3. Klemenjak, Christoph and Elmenreich, Wilfried. On the Applicability of Correlation Filters for Appliance Detection in Smart Meter Readings. In Proceedings of the 2017 IEEE International Conference on Smart Grid Communications (SmartGridComm), Dresden, Germany, 2017. 4. Kim, Jihyun, Thi-Thu-Huong Le, and Howon Kim. "Nonintrusive Load Monitoring Based on Advanced Deep Learning and Novel Signature." Computational intelligence and neuroscience 2017 (2017). 5. Faustine, Anthony, et al. "A Survey on Non-Intrusive Load Monitoring Methodies and Techniques for Energy Disaggregation Problem." arXiv preprint arXiv:1703.00785 (2017). 6. Zhang, Chaoyun, et al. "Sequence-to-point learning with neural networks for nonintrusive load monitoring." arXiv preprint arXiv:1612.09106 (2016). 7. Klemenjak, Christoph, Dominik Egarter, and Wilfried Elmenreich. "YoMo: the Arduino-based smart metering board." Computer Science-Research and Development 31.1-2 (2016): 97-103. 8. Klemenjak, Christoph, and Peter Goldsborough. "Non-Intrusive Load Monitoring: A Review and Outlook." arXiv preprint arXiv:1610.01191 (2016). 9. Kelly, Jack, and William Knottenbelt. "Neural NILM: Deep neural networks applied to energy disaggregation." Proceedings of the 2nd ACM International Conference on Embedded Systems for Energy-Efficient Built Environments. ACM, 2015. 10. Egarter, Dominik, and Wilfried Elmenreich. "Load disaggregation with metaheuristic optimization." Energieinformatik, Karlsruhe, Germany (2015). 11. Armel, K. Carrie, et al. "Is disaggregation the holy grail of energy efficiency? The case of electricity." Energy Policy 52 (2013): 213-234. 12. Zoha, Ahmed, et al. “Non-intrusive load monitoring approaches for disaggregated energy sensing: A survey.” Sensors 12.12 (2012): 16838-16866.
  26. 26. Questions? https://about.me/klemenjak
  27. 27. Limitations ● Feedback is helpful but is it really worth the effort of NILM? ● Smart home integration ● Use cases beyond energy insights ● Less frequent but more meaningful user engagement ● Open independent performance assessments ● Common industry definition of features ● Standard industry API definitions Inspired by Will Siddall: Why isn’t NILM mainstream yet? A view from both sides. Talk at EU NILM 2017

×