Applying machine learning techniques for forecasting flexibility of virtual power plants

Pamela MacDougall, Anna Magdalena Kosek, Henrik W. Bindner, Geert Deconinck

Research output: Chapter in Book/Report/Conference proceedingArticle in proceedingsResearchpeer-review

Abstract

Previous and existing evaluations of available flexibility using small device demand response have typically been done with detailed information of end-user systems. With these large numbers, having lower level information has both privacy and computational limitations. We propose a black box approach to investigating the longevity of aggregated response of a virtual power plant using historic bidding and aggregated behaviour with machine learning techniques. The two supervised machine learning techniques investigated and compared in this paper are, multivariate linear regression and single hidden layer artificial neural network (ANN). Both techniques are used to model a relationship between the aggregator portfolio state and requested ramp power to the longevity of the delivered flexibility. Using validated individual household models, a smart controlled aggregated virtual power plant is simulated. A hierarchical market-based supply-demand matching control mechanism is used to steer the heating devices in the virtual power plant. For both the training and validation set of clusters, a random number of households, between 200 and 2000, is generated with day ahead profile scaled accordingly. Further, a ramp power (power deviation) is assigned at various hours of the day and requested to hold for the remainder of the day. Using only the bidding functions and the requested ramp powers, the ramp longevity is estimated for a number of different cluster setups for both the artificial neural network as well as the multi-variant linear regression. It is found that it is possible to estimate the longevity of flexibility with machine learning. The linear regression algorithm is, on average, able to estimate the longevity with a 15% error. However, there was a significant improvement with the ANN algorithm achieving, on average, a 5.3% error. This is lowered 2.4% when learning for the same virtual power plant. With this information it would be possible to accurately offer residential VPP flexibility for market operations to safely avoid causing further imbalances and financial penalties.
Original languageEnglish
Title of host publicationProceedings of 2016 IEEE Electrical Power and Energy Conference
Number of pages6
PublisherIEEE
Publication date2016
Pages1-6
ISBN (Print)9781509019199
DOIs
Publication statusPublished - 2016
EventIEEE Electrical Power and Energy Conference 2016 - Ottawa, Canada
Duration: 12 Oct 201614 Oct 2016

Conference

ConferenceIEEE Electrical Power and Energy Conference 2016
CountryCanada
CityOttawa
Period12/10/201614/10/2016
Series2016 Ieee Electrical Power and Energy Conference (epec)

Keywords

  • Space heating
  • Water heating
  • Heat pumps
  • Linear regression
  • Mathematical model
  • Power generation
  • Load management
  • Demand response
  • Neural Networks
  • Smart Grids
  • Heating Systems
  • Prediction
  • Energy flexibility
  • Aggregation

Fingerprint Dive into the research topics of 'Applying machine learning techniques for forecasting flexibility of virtual power plants'. Together they form a unique fingerprint.

Cite this