Effect of short data periods on the annual prediction accuracy of temperature-dependent regression models of commercial building energy use

Ideally, a full year or more of energy use and weather data should be used to construct empirical models of building energy use. However, in many cases a full year of data is not available and one is constrained to develop models using less than a full year of data. This paper examines how temperature dependent regression models of energy use based on periods of less than one year fare in terms of annual predictive ability compared to models based on a full year of data. The primary methodology employed is to construct temperature-dependent linear regression models of daily energy use from one, three and five month data-sets and compare the annual energy use predicted by these models to the annual energy use predicted by a model based on an entire year of data. Heating and cooling energy use from the buildings in Central Texas were examined. On average, the annual energy loss predicted by chilled water use models based on three month data-sets varied by about 4%, and never varied by more than 20% from the annual energy use even when data-sets as short a one month were used. The values predicted by the heating energy use models, on the other hand, varied by as much as 400% from the annual energy use, though on average the errors were around 20% for models based on three months of data. Several characteristics of the data-sets and models which influenced their predictive ability were identified, The most important being that a short data set's annual predictive ability is strongly influenced by the proximity of the short data set's mean temperature to the annual mean temperature.