Automated and comprehensive day-type grouping in building energy baseline models with a series of statistical inferences

Abstract

A good model fit is the central pillar of building energy consumption baseline modeling. Energy analysts often must undertake the arduous and tedious task of separating measured energy consumption data into groups reflecting a building’s multiple operation modes. While a building sometimes has more than a weekday/weekend shift, a detailed operation schedule may not be easily available or accurate. If not treated carefully, operational idiosyncrasies can limit the performance of the baseline model. So far, the data separation methods deployed in published articles are either too simple to be effective for complex schedules or so intricate that many working in the field may find them inconvenient to implement. The proposed procedure in this study instead adopts a novel approach that automatically and comprehensively examines all data separation possibilities with pre-defined elementary day-types through a series of lack-of-fit F-tests. It then suggests a best one that balances model accuracy and simplicity. This procedure was tested on measured energy consumption data of 76 buildings and found to weigh simplicity more heavily than traditional statistical complexity-penalising metrics. It improved the CV-RMSE by 7.6 percentage points on average and helped extract information from the data to understand better the buildings’ energy consumption patterns.