The default Time-Of-Week and Temperature model allows for extended baseline periods when fitting baseline models. When the model adaptation function is not used, a single model can be fit to the entire baseline period, which could be up to 12 months long. The single, yearly regression approach assumes that base load and weather sensitivity of energy consumption is constant throughout the year.
Empirical evidence shows that baseline and weather-related energy use varies during different months of the year. This variation is not represented when a single regression is estimated for the entire baseline period. Below are two potential modeling approaches:
- A regression approach that estimates one model for the entire baseline period, which is 1 year in this case.
- A regression approach that estimates 12 models for each month of the baseline period, which is 1 year in this case.
After reviewing the results of the empirical testing, we recommend applying a three-month weighted regression model for residential hourly methods. Twelve models will be fit for each month of the year, with months before and after the month of interest weighted down by 50%. For example, when predicting the counterfactual energy usage for the month of July, the corresponding baseline model will be fit using data from June, July and August of the previous year. The data points from June and August will be assigned a 50% weight compared to the data points from July. This approach accounts for varying energy consumption patterns across months of the reporting period without overfitting the model to limited data.
Keep an eye out for next week’s blog post where we’ll summarize the testing of valuation methods for hourly savings.