Identifying critical building-oriented features in city-block-level building energy consumption: A data-driven machine learning approach

Abstract Understanding regional building energy patterns is the prerequisite to efficiently and effectively promote sustainable urban development. Previous studies have proposed various data-driven methods to investigate the relationship between building energy consumption and hundreds of potential influencing features. However, it is difficult to include all potential features in one single model since either some data could be unavailable or the model would be too complex. To identify the critical features, this study develops a data-driven random forest (RF) based framework with a dataset of Taipei City, consisting of 24,764 buildings in 881 city-blocks, to model the relationship between city-block-level building-oriented features and building energy consumption. The RF model is found to outperform other machine learning models including logistic regression, k-nearest neighborhood, support vector machine, and decision tree models in the predictive accuracy of the classification problem. Seven critical features related to the built year of buildings, building gross floor area, building density, and the ratio of commercial buildings in the block are identified from the 59 city-block-level building-oriented features. The developed framework could refine the features adopted in regional building energy models, and policymakers and city planners would get practical implications from the identified critical features.