CSE, with support from the University of Bristol, worked on a two year project to create a new tool to estimate the thermal efficiency of a home based on energy consumption data and temperature data.

The project was funded by the Department for Business, Energy and Industrial Strategy (BEIS) through the Smart Meter Enabled Thermal Efficiency Ratings (SMETER) Innovation Competition.

Many of the policies to reduce carbon emissions and energy consumption in homes are based on having an understanding of their energy performance. Currently, the main method used by Government to compare and assess the energy performance of homes is the Standard Assessment Procedure (SAP).

However, there is evidence to suggest that the variable quality of survey data used in SAP calculation adversely impacts the accuracy.

Improving the accuracy of SAP calculations with machine learning

To improve accuracy, CSE’s SMETER project aimed to improve the part of the calculation procedure that estimates the thermal performance of the building envelope – termed the ‘heat transfer coefficient’ or ‘HTC’- by basing the calculation on real, measured data.

At present, the best available method to measure this thermal performance is the co-heating test. But this is costly to carry out and requires the home to be empty for a two-week period, making it unsuitable for widespread use.

CSE tested the use of machine learning methods to create a methodology that can be deployed less intrusively and more cost-effectively.

A more accurate and dwelling-specific heat transfer coefficient could also have uses outside of policy-making and the SAP calculation process. CSE are particularly interested in its potential to help identify and support people living in cold homes.

An assessment team composed of researchers from UCL and the University of Loughborough tested our methodology, alongside those of other competition participants, including a trial in real homes facilitated by Halton Housing.

Whilst the initial results were promising, it was found to be difficult to access smart meter data and temperature data at sufficient scale to effectively train our model and to have sufficient confidence in its outputs. The code and supporting documentation from this project are available open source to any interested parties.