Abstract

Interpretable and scalable data-driven methodologies providing high granularity baseline predictions of energy use in buildings are essential for the accurate measurement and verification of energy renovation projects and have the potential of unlocking considerable investments in energy efficiency worldwide. Bayesian methodologies have been demonstrated to hold great potential for energy baseline modelling, by providing richer and more valuable information using intuitive mathematics. This paper proposes a Bayesian linear regression methodology for hourly baseline energy consumption predictions in commercial buildings. The methodology also enables a detailed characterization of the analyzed buildings through the detection of typical electricity usage profiles and the estimation of the weather dependence. The effects of different Bayesian model specifications were tested, including the use of different prior distributions, predictor variables, posterior estimation techniques, and the implementation of multilevel regression. The approach was tested on an open dataset containing two years of electricity meter readings at an hourly frequency for 1578 non-residential buildings. The best performing model specifications were identified, among the ones tested. The results show that the methodology developed is able to provide accurate high granularity baseline predictions, while also being intuitive and explainable. The building consumption characterization provides actionable information that can be used by energy managers to improve the performance of the analyzed facilities.

Abstract

Increasing the energy efficiency of the built environment has become a priority worldwide and especially in Europe. Because of the relatively low turnover rate of the existing built environment, energy efficiency retrofitting appears to be a fundamental step in reducing its energy consumption. Last experiences have shown that there is a vast energy efficiency potential lying in the building stock, and it is mainly untapped. One of the reasons is a lack of robust methodologies able to evaluate the effect of applied energy efficiency measures and inform about the expected impact of potential retrofitting strategies. Nowadays, dynamic measured data coming from automated metering infrastructure provides valuable information to evaluate the effect of energy conservation strategies. For this reason, energy performance modeling and assessment methods based on this data are starting to play a major role. In this paper, several methodologies for the measurement and verification of energy savings, and for the prediction and recommendation of energy retrofitting strategies, are analysed in detail. Practitioners looking at different options for these two processes, will find in this review a thorough and detailed overview of the different methods that can be used. Guidance is also provided to determine which method could work best depending on the specific case under analysis. The reviewed approaches include statistical learning models, machine learning models, Bayesian methods, deterministic approaches, and hybrid techniques that combine deterministic and data-driven modeling. Existing research gaps are identified and prospects for future investigation are presented within the main conclusions of this research work.