Bridging the gap between data-driven baselines and energy saving uncertainty for building retrofit

Abstract

Data-driven models are increasingly used as baselines for evaluating energy savings from building retrofit measures. However, a critical challenge arises because the inherent model prediction error of these data-driven models is frequently comparable to the measured energy saving percentage itself. The impact of inherent model prediction errors and uncertainties on the reliability of energy saving estimates has often been overlooked. This study proposes a simple and statistical framework that establishes a reliable, quantitative relationship between the trusted energy saving percentage and three readily available parameters: the model error level (measured by CVRMSE), the volume of post-retrofit data, and the variability of the predicted baselines. The findings demonstrate that, under a given significance level, the trusted energy saving is the observed energy difference penalized by a negative term representing an uncertainty penalty. This study analytically shows that this penalty is magnified by higher model error and greater variability in predicted baselines, but is effectively reduced by a larger volume of post-retrofit data. The resulting framework provides a direct formula to quantify the confidence level of saving estimates, offering a clearer understanding of the confidence associated with energy efficiency investments.