Interpretable domain-informed and domain-agnostic features for supervised and unsupervised learning on building energy demand data

Energy demand from the built environment is among the most important contributors to greenhouse gas emissions. One promising way to curtail these emissions is through innovative energy management systems (EMS's). These systems often rely on access to real-world demand data, which remains elusive in practice. Even when available, energy demand data typically suffers from missing data as well as many irregularities and anomalies. This precludes the application of many off-the-shelf machine learning algorithms for time series analysis and modelling, necessary for downstream energy management. Transforming energy demand time series to low dimensional feature matrices has been shown to work well in determining similar buildings and predicting meta-data, both of which can be used to create better forecast algorithms used as input in EMS's. These studies are, however, often marred by the limited size of datasets, as well as the non-interpretable nature of extracted features. This paper addresses these concerns and makes several important contributions: (1) it collates several open-source datasets to create a large meta-analysis dataset containing energy demand data for over 13,000 buildings; (2) it investigates the use of different interpretable feature extraction methods on this collated dataset; and (3) it shows that this feature matrix can be used more generally to determine similar buildings and predict building properties such as missing meta-data. The large feature matrix resulting from the work is open-sourced as part of a web-based dashboard to enable the community to reproduce and further develop our results.