Though many building energy benchmarking programs have been developed during the past decades, they hold certain limitations. The major concern is that they may cause misleading benchmarking due to not fully considering the impacts of the multiple features of buildings on energy performance. The existing methods classify buildings according to only one of many features of buildings the use type, which may result in a comparison between two buildings that are tremendously different in other features and not properly comparable as a result. This paper aims to tackle this challenge by proposing a new methodology based on the clustering concept. The clustering concept, which reflects on machine learning algorithms, classifies buildings based on a multi-dimensional domain of building features, rather than the single dimension of use type. Buildings with the greatest similarity of features that influence energy performance are classified into the same cluster, and benchmarked according to the centroid reference of the cluster. The proposed methodology contains four steps: feature selection, clustering algorithm adaptation, results validation, and interpretation. The experimentation was carried out with a comparison between the proposed methodology and the Energy Star approach. It was shown that the proposed methodology could account for the total building energy performance and was able to provide a more comprehensive approach to benchmarking. In addition, the multi-dimensional clustering concept enables energy benchmarking among different types of buildings, and inspires a new perspective to investigate building performance typology. (C) 2014 Elsevier B.V. All rights reserved.