E-Adivino: A Novel Framework for Electricity Consumption Prediction based on Historical Trends

Shubham Saini, Pandarasamy Arjunan, Amarjeet Singh, Ullas Nambiar


Abstract:

Prediction of electricity demand is becoming increasingly important for several real world applications. One such application, used by electrical utilities across the world, is the Demand Response (DR) program for peak demand management. For a real world utility having a large number of customers, learning the best fit baseline projection instance for every consumer may be time consuming. This is further complicated by fluctuating demand across different time periods thus possibly requiring multiple baseline projection instances. Motivated by such a requirement, we propose E-Adivino: an electricity forecasting framework that first clusters consumption pattern using temporal characteristics followed by forecasting for each cluster using a generalized baseline projection framework. E-Adivino allows for selection and application of appropriate forecasting models for different consumers based on their demand patterns rather than applying a uniform model for all the consumers, as is the practice today. E-Adivino is evaluated for its real world applicability using consumption pattern of 20 different loads collected from a university campus in India over a duration of six months along with year long consumption data from commercial consumers in the USA, as provided by EnerNOC.

Contributions of this work:
  • E-Adivino forecasting framework that includes a generalized baseline projection framework exploiting lookback window, exclusion rules and forecasting models.
  • Time series clustering within the E-Adivino framework, that uses different temporal characteristics of the consumption pattern to group multiple loads together so as to reduce the computation effort for learning appropriate baseline projection models for a large set of loads.
  • Validation of the proposed framework using real word dataset spanning across two different countries - loads from IIIT Delhi campus in India, as collected by us and a set of commercial consumers in the USA, as released by EnerNOC.
  • We release our IIIT Delhi dataset of 20 loads, used for evaluation in this paper, with power consumption at 5 minute granularity for a duration of 6 months.
IIIT Delhi Campus Dataset

Building NamePricing TypeNo. of FloorsArea Covered (sqm)Daily Consumption (KWh)
Academic BuildingCommercial61000700
LibraryCommercial4620225
FacilitiesCommercial3577250
Mess and ActivityCommercial41279450
Boys HostelResidential81116275
Girls HostelResidential5838150
Faculty HousingResidential12560700

Electricity consumption data for 20 highest consumption commercial loads. The data spans across 6 months durations (from 1st January 2014 to 30th June 2014) and each value is an average over 5 minutes time interval. Metadata includes name of appliance/load, name of building, pricing type, meter make, meter model and meter id.