Household data analytics Dagstuhl, February 2015 Christoph - - PowerPoint PPT Presentation

household data analytics
SMART_READER_LITE
LIVE PREVIEW

Household data analytics Dagstuhl, February 2015 Christoph - - PowerPoint PPT Presentation

Technische Universitt Mnchen Household data analytics Dagstuhl, February 2015 Christoph Doblander, Anwar Ul Haq, Christoph Goebel, H.-A. Jacobsen Technische Universitt Mnchen Institut fr Informatik, Lehrstuhl I13 Prof. Dr. H.-A.


slide-1
SLIDE 1

Technische Universität München

Household data analytics

Dagstuhl, February 2015

Christoph Doblander, Anwar Ul Haq, Christoph Goebel, H.-A. Jacobsen

Technische Universität München Institut für Informatik, Lehrstuhl I13

  • Prof. Dr. H.-A. Jacobsen

Alexander von Humboldt-Professor

slide-2
SLIDE 2

Technische Universität München

Peer Energy Cloud

Department of Computer Science, Chair for Application & Middleware Systems

2

slide-3
SLIDE 3

Technische Universität München

Our contributions within the project

  • Energy analytics framework

– Highly scalable framework which uses spare compute resources

  • Evaluation of forecasting methods

– Algorithms: SVM, naive bayes, ... – Benchmark: persistence, average load profiles and same as last weekend forecast – Timehorizon: {15, 30, 60, 90 } ahead minutes horizon

  • Not shown today

– Anomaly detection – Model based forecasting – Device detection

Department of Computer Science, Chair for Application & Middleware Systems

3

slide-4
SLIDE 4

Technische Universität München

Relevance of prediction of a single household

  • Increase self sufficiency is a financial benefit
  • Hard to justify here in Germany because of high

incentives to feed electricity into the grid

Department of Computer Science, Chair for Application & Middleware Systems

4 Example Austria/Tirol 5kWp cost roughly 9,000 € 1,375 € subsidiary once for investing Feeding into the grid When lucky, OEMAG quota (12 ct/Kwh, maxed out in the first few hours) else contract with a electricity trader between 2.4 – 5 ct/Kwh compared to 12 ct/Kwh in Germany

New Quota 07.01.2015 17:00

slide-5
SLIDE 5

Technische Universität München

Department of Computer Science, Chair for Application & Middleware Systems

5

No joke: http://www.unijobs.at/nebenjobs/vorschau/248788

slide-6
SLIDE 6

Technische Universität München

Current possibilities

  • Miele@Home

– „SG-ready“ – Miele Smart Start

  • Extension

– Predict non-controllable loads, schedule the others accordingly – One input of the control is the prediction – => Increase self sufficiency

Department of Computer Science, Chair for Application & Middleware Systems

6

slide-7
SLIDE 7

Technische Universität München

Windowing

Department of Computer Science, Chair for Application & Middleware Systems

7 Prediction Base Actuals

Interval Forecast Minutes Backlog Minutes

slide-8
SLIDE 8

Technische Universität München

Machine learning pipeline

Extraction

  • Raw CSV files are extracted line by line

Windowing

  • Learn horizon
  • Prediction horizon
  • Every increment, e.g., 30 minutes

Prediction

  • Results per window

Error calculation

  • MAPE
  • RMSE
  • ...

Department of Computer Science, Chair for Application & Middleware Systems

8

Map Fold or Map Map Reduce

slide-9
SLIDE 9

Technische Universität München

Feature extraction

  • Statistical

– Min, Max, Mean, Variance, Standard deviation

  • Descriptive statistics

– Skewness, Kurtosis

  • Financial

– Momentum, WilliamsR

  • Discrete

– Last state on/off – Switched on/off

  • Timestamp extraction

– Week day – Hour of the day

Department of Computer Science, Chair for Application & Middleware Systems

9

slide-10
SLIDE 10

Technische Universität München

Department of Computer Science, Chair for Application & Middleware Systems

10

slide-11
SLIDE 11

Technische Universität München

Results

Department of Computer Science, Chair for Application & Middleware Systems

11

slide-12
SLIDE 12

Technische Universität München

Results overview and conclusion

  • 34 households

– 20 households were chosen for more evaluation – 14 households were rejected because of bad data quality – Benchmark is persistence – Results are preliminary – 70% of the data analysed

  • Benchmark can be beaten

– But prediction error is still high

  • More features tend to reduce predictionerror in

predictions up to an hour

Department of Computer Science, Chair for Application & Middleware Systems

12

slide-13
SLIDE 13

Technische Universität München

Future work

  • Many problems, …
  • But it’s Opensource 
  • Already done
  • Own software for getting the results
  • Own backend
  • Current stage
  • Gain experience

Department of Computer Science, Chair for Application & Middleware Systems

Energy Monitoring Node

slide-14
SLIDE 14

Technische Universität München

References

  • A. Veit, C. Goebel, R. Tidke, C. Doblander, H.-A.
  • Jacobsen. Household Electricity Demand Forecasting:

Benchmarking State-of-the-Art Methods. 5th ACM International Conference on Future Energy Systems (ACM e-Energy), Cambridge, UK.

  • H. Ziekow, C. Doblander, C. Goebel, H.-A. Jacobsen. Forecasting

Household Electricity Demand with Complex Event Processing: Insights from a Prototypical Solution. 13th ACM International Middleware Conference, Beijing, China. 2013

  • H. Ziekow, C. Goebel, J. Strüker, H.-A. Jacobsen. The Potential of

Smart Home Sensors in Forecasting Household Electricity

  • Demand. IEEE International Conference on Smart Grid

Communications (SmartGridComm2013), Vancouver, Canada. 2013

Department of Computer Science, Chair for Application & Middleware Systems

14