What's in the box?! Towards explainable machine learning applied to non-residential building smart meter classification

Feature engineering and data-driven classification models are at the forefront of analysis of large temporal sensor data from the built environment. In previous efforts, temporal features were engineered from the whole building hourly electrical meter data from 507 non-residential buildings. These features fall within the three general categories of statistics, model, and pattern-based and can be used to identify various behavior in the structure of the whole building electrical meter data. In this paper, a deeper investigation is made of exactly what types of behavior are most important in the context of two classification scenarios: the primary use of a building and the level of performance the building has when compared to its peers. The highly comparative time-series analysis (hctsa) toolkit is used to analyze the most important temporal features for the classification of various building performance attributes. In the first analysis, a comparison is made to distinguish the behavior between university dormitories (70 buildings) and laboratories (95 buildings) as an example of interpreting the classification of the primary-use-type of a building. In the second analysis, a comparison of buildings with high (165 buildings) versus low (169 buildings) consumption is used to extract and understand the behavior that indicates the level of the energy performance of a building. These two case study examples provide a foundation for further explainable machine learning techniques in both classification and prediction as applied to buildings. This effort is the first example of machine learning with an explicit focus on the interpretability of classification for smart meter data from non-residential buildings. (C) 2019 Elsevier B.V. All rights reserved.