Machine learning applications in urban building energy performance forecasting: A systematic review

Abstract

Abstract In developed countries, buildings are involved in almost 50% of total energy use and 30% of global green-house gas emissions. Buildings' operational energy is highly dependent on various building physical, operational, and functional characteristics, as well as meteorological and temporal properties. Besides physics-based building energy modeling, machine learning techniques can provide faster and higher accuracy estimates, given buildings' historic energy consumption data. Looking beyond individual building levels, forecasting buildings’ energy performance helps city and community managers have a better understanding of their future energy needs, and plan for satisfying them more efficiently. Focusing on an urban-scale, this study systematically reviews 70 journal articles, published in the field of building energy performance forecasting between 2015 and 2018. The recent literature have been categorized according to five criteria: 1. Learning Method, 2. Building Type, 3. Energy Type, 4. Input Data, and 5. Time-scale. The scarcity of building energy performance forecasting studies in urban-scale versus individual level is considerable. There is no study incorporating building functionality in terms of space functionality share percentages, nor assessing the effects of climate change on urban buildings energy performance using machine learning approaches and future weather scenarios. There is no optimal criteria combination for achieving the most accurate machine learning-based forecast, as there is no universal measure able to provide such global comparison. Accuracy levels are highly correlated with the characteristics of forecasting problems. The goal is to provide a comprehensive status of machine learning applications in urban building energy performance forecasting, during 2015–2018.