Classification of retrofit measures for residential buildings according to the global cost

Abstract

Data-driven black-box surrogate models are widely used in research related to buildings energy efficiency. They are based on machine learning techniques, learn from available data, and act as a replacement for or an addition to complex and computationally intensive knowledge-based models. Surrogate models can predict energy demand, indoor air temperature, or occupants behavior; explore search space in optimization problems; learn control rules; etc. This paper analyzes surrogate models that classify building retrofit measures directly according to the global cost. In addition, they quantify the importance of each variable for the classification process. The models are based on random forest classifiers, which are fast and powerful ensemble learners. They can be applied to effectively reduce search spaces when optimizing energy renovation measures or to rapidly identify projects that deserve financial support. This approach is applied to two residential buildings and three scenarios of price development. The training process uses a small share of retrofit options assessed with standard calculations of the heating and cooling demands, as well as the global cost. The results show very high classification performance, even when the models are trained with small and imbalanced training sets. The obtained recall, precision, and F-score values are mostly above 95%, except for extremely small training sets.