Data-driven baseline generation for post-retrofit energy saving assessment, a comparison of statistical and machine learning methods

The renovation wave aims to improve energy performance of the existing building stock, encouraging development of methods for predicting post-retrofit energy use and quantifying savings. Therefore, establishing an energy use baseline is crucial for computing changes and savings. Data-driven techniques vary in effectiveness for this purpose. Selecting a suitable data-driven method for post-retrofit energy use modelling requires choosing between several approaches. This study compares two methods of establishing a baseline for post-retrofit evaluation, focusing on predicting heating and electricity energy use. One method utilizes monthly degree-day normalisation and baseline derivation, and the other employs machine learning techniques, including clustering to address seasonal variations and non-linear regression models like random forest and neural network. The comparison was based on CVRMSE, obtained by applying the methods to eight datasets. These individual datasets included metered and simulated data for heating and electricity energy use in two large non-residential buildings. An upper error margin of CVRMSE 25 % for annual energy use, as set in ASHRAE Guidelines, was not reached. The actual uncertainty during validation with simulated data ranged from 4.5 to 10.4 % for heating and electricity models using the degree-day method, and from 1.4 to 7.5 % for machine learning models. When applied to metered data, the degree-day method showed an uncertainty of 8.4–19.3 %, while machine learning models had an uncertainty of 6.11–17.5 %. Additionally, monthly percentage error analysis confirmed a considerably better performance of the machine learning models. This study contributes to the assessment of renovation impact and operational energy savings by offering additional perspectives on selecting energy use modelling methods, which are also applicable in the context of Minimum Energy Performance Standards.