A learning-based time-efficient framework for building energy performance evaluation

Abstract

Recently, detailed building energy models are being widely used for building control assessments. When conducting such assessments, year-long simulations are usually deployed to cover as many building operation conditions as possible. However, year-long simulations can become computationally cumbersome, given the inherent nonlinearities of high-fidelity, detailed building models. In this paper, we propose a novel, time-efficient, and data-driven framework for selecting specific weather conditions when assessing building energy performance within a simulation environment. In the proposed framework, learning-based methodologies are used to group days of a performance assessment time window (usually a whole year) which are “similar” in terms of weather conditions (such as temperature, humidity, precipitation etc.) or contexts (such as “day of the week”, “weekday/weekend”, etc.) into different clusters. Following this, we select a finite number of “representative days” from each of these clusters, using which we try to understand the correlations between building performance and weather (or contexts) for each specific cluster. Subsequently, this knowledge is used to predict the building performance for the remaining “non-representative days”, and in turn estimate the overall annual energy performance of the whole building. We demonstrate the performance of the proposed framework on a varied suite of buildings across different geographical locations of the U.S. Results suggest that for practical purposes, our proposed approach can reduce the computational cost while maintaining considerable accuracy, compared to year-long simulations.