Inferring building type using textual data and Natural Language Processing for urban building energy modelling

Abstract

Building type is among the most important inputs for building energy model. However, the information of building type is always missing in urban scale building energy modeling. This paper presents a novel approach to infer building type from building name. First, we created the building name text dataset through the fusion of GIS spatial data. A rule-based method was developed to estimate building types using naming features. We then trained five machine learning classifiers, including four transformer models and one Multilayer Perceptron model, to predict building types. Finally, we leveraged the inferred building type information for building energy consumption simulation, addressing the crucial data scarcity issue in urban-scale building energy models. Experimental results indicated that our rule-based classification method achieved a precision of 84.3%. The RoBERTa model, the best-performing natural language processing (NLP) model, reached a precision of 91.6% with both Chinese and English names as NLP model inputs, showcasing a 1.3% enhancement compared to solely utilizing the Chinese dataset and a 1.8% improvement compared to solely utilizing the English dataset. This research proposes a useful framework to infer building type by leveraging the state-of-art NLP techniques, paving the way for more accurate and efficient urban-scale building energy modelling.