Clustering and Visualisation of Electricity Data to Identify Demand Response Opportunities

A poster-abstract data-driven method that clusters and visualises raw building electricity data to produce day-type profiles – winter weekdays, weekends/holidays, and seasonal variants. Pre-processing handles summer/winter timestamp artefacts, multivariate outlier detection cleans the data, and k-medoids on 24-feature daily vectors with silhouette-width selection produces the clusters. Used as a screening step to rank a portfolio of retail sites by ‘how worth-auditing they look from meter data alone’ in the FlexReStore project.

research
clustering
demand response
electricity data
retail
Poster Abstract · 3rd ACM International Conference on Systems for Energy-Efficient Built Environments (BuildSys 2016), pp. 233–234.
Authors

Almir Mehanovic

Emil Sebastian Rømer

Jakob Hviid

Mikkel Baun Kjærgaard

Published

November 16, 2016

Publication

Abstract

Electricity grids are facing challenges due to peak consumption and renewable electricity generation. In this context, demand response offers a solution to many of the challenges, by enabling the integration of consumer side flexibility in grid management. Retail buildings are good candidates for providing flexible demand due to their volume and the stability of their loads.

However, new methods are needed to efficiently identify demand response opportunities in retail buildings. In this poster we outline a data-driven method based on clustering and visualisation that generates day-type profiles from raw electricity consumption data. The day-type profiles among others enable analysis of the repeatability and seasonal variation of building loads. Proposing such a method is a step towards enabling a higher penetration of intelligent smart grid solutions in the retail sector.

Method

The method runs in three steps:

  1. Pre-processing – normalise timestamps to handle summer/winter offsets and similar artefacts.
  2. Outlier detection – remove extreme outliers using a multivariate algorithm before clustering.
  3. Clustering – k-medoids on 24-feature daily vectors plus statistical extensions, with optimal cluster count selected by silhouette width.

Day-type profiles emerge for, e.g., winter weekdays, weekends/holidays, and seasonal variants – each then a candidate for DR audit prioritisation.

Why this matters

Site audits are expensive. A method that can rank buildings by how worth-auditing they look from their meter data alone turns audits from “screen everyone exhaustively” into “screen the buildings the data already flags as interesting.” In the FlexReStore project this was applied to a portfolio of retail sites.

Citation

Mehanovic, A., Rømer, E. S., Hviid, J., & Kjærgaard, M. B. (2016). Clustering and Visualisation of Electricity Data to identify Demand Response Opportunities: Poster Abstract. In Proceedings of the 3rd ACM International Conference on Systems for Energy-Efficient Built Environments, pp. 233–234. ACM. https://doi.org/10.1145/2993422.2996403