Skip to main content
Using Pattern Recognition to Communicate the Performance Gap
Back To All Knowledge Items

Using Pattern Recognition to Communicate the Performance Gap

Standard Rate
Member Rate

Session 9, Paper 1, CIBSE ASHRAE Technical Symposium, Dublin, Ireland, 3-4 April 2014


The gap between calculated and actual energy use in buildings (the energy performance gap) is now a well-known problem, and has been highlighted and quantified in part by the data collected in the CIBSE/RIBA CarbonBuzz project. In the data gathered to date, the median energy performance gap, defined as the ratio between calculated and measured performance, is 1.9 for schools and 2.0 for offices.
This paper presents an exploratory analysis of the data. It investigates the potential of nearest-neighbour regression to identify the group of records which are most similar to a given target building and use that group to estimate the size of the performance gap in the target building. The crowd-sourced nature of the data set means that it is incomplete for all records, necessitating significant pre-processing of the data. In particular the substitution of calculated values for missing data and automated elimination of outliers are employed.
A number of model parameters are considered including selection of features to include in the model, and the size of the group of neighbours. Results are compared against a linear regression model, TM46 benchmark predictions and the mean performance gap from the data set. The kNN method is found to be the best predictor of the performance gap, though the predictions exhibit high variance and low accuracy. Calculated energy consumption and carbon emissions per m2, and building type (whether existing, refurbished or new build) are found to be the most predictive features.
Finally, work is presented on the CarbonBuzz Metadata project, funded by the Technology Strategy Board, including a proposed interface for finding and viewing similar buildings. This is intended to communicate the possible energy performance gap based on that found in similar buildings. The approach demonstrated employs an algorithm based on the work undertaken in this study.