The CalTRACK Hourly Model is used when hourly or sub-hourly consumption data is available. Sub-hourly data is aggregated to the hourly level before using the model. CalTRACK recommends the use of the Time Of Week and Temperature (TOWT) model developed by Lawrence Berkeley National Labs (LBNL) with standardization of certain user-defined inputs and a different treatment of the baseline to allow for longer term load predictions.
The hourly model explains a building energy use as the interaction between the building’s temperature dependence, the occupancy status and the time of week. Most buildings tend to have relatively predictable weekly schedules. In addition, the outdoor temperature is generally a large driver of the HVAC systems that consume a large portion of the building’s energy. However, the impact of outdoor temperature on energy consumption can be different, depending on whether the building is occupied or unoccupied. High occupancy usually corresponds to higher energy use.
Three types of variables are generated before fitting a CalTRACK hourly model - temperature, time-of-week and occupancy.
For each temperature data point in the baseline dataset, the outdoor air temperature is used to calculate up to 7 new binned features using an algorithm developed for the original TOWT model. This algorithm apportions the temperature into different bins with known endpoints as shown in the figure below.
A week is divided into 168 hourly time-of-week intervals starting on Monday. For example, interval 1 is from midnight to 1 a.m. on Monday morning, interval 2 is from 1 a.m.-2 a.m. and so on. Dummy variables are included in the model for each time of week.
The sensitivity of building energy use to temperature may vary depending on the “occupancy” status. This is handled in the Time-Of-Week and Temperature model by segmenting the times-of-week into periods of high load and low load (also referred to as occupied/unoccupied, although the states may not necessarily correspond to occupancy changes). The segmentation is accomplished using the residuals of a HDD-CDD model that uses fixed balance points. Hours of the week that appear to be in high usage mode most of the time are flagged as “occupied”.
The model specification is then applied using these variables - refer to the CalTRACK documentation for full technical specifications.
CalTRACK tests revealed that using a single baseline model to model a building’s energy consumption yielded models that were unbiased over the whole year, but tended to have strong seasonal biases which affected energy savings results. To counter this issue, instead of using a single baseline model for estimating the counterfactual during all times of the year, predicting the counterfactual during any time period is done using the baseline model for that calendar month (“month-by-month” models). This implies that there can be up to 12 separate models for a particular building - one for predicting the counterfactual in each calendar month. Each model is fit using baseline data comprising (i) data from the same calendar month in the 365 days prior to the intervention date. These data points are given full weight when fitting the model, (ii) data from the previous and subsequent calendar months in the 365 days prior to the intervention date. These data points are given a weight of 0.5 when fitting the model. For example, for a project installed in March 2018, predicting the counterfactual in July 2018 will be done using a model fit to baseline data from June, July and August 2017, with weights of 0.5, 1 and 0.5 assigned to the data points in those three months.
Phil price report
Mathieu et al paper