Comparison of data-driven building energy use models for retrofit impact evaluation

A change-point (piecewise linear regression) model fitted to the pre-retrofit data as the counterfactual for the savings calculation, is considered to be the best approach to evaluating the energy savings of building retrofits ( ASHRAE Guideline 14). However, when applied to a large portfolio savings analysis with substantial multi-year data, the change-point model does not fit the data well in some cases. The study thus aims to improve the accuracy of the changepoint model by: 1) using more advanced non-linear models, 2) incorporating additional input features, and 3) increasing the time resolution of input variables. We found that random forest regression (RF) models with an array of climate (humidity, wind, solar radiation, etc.), time (day of the week, season, holiday), and energy consumption of the immediate past 1-4 hours (energy lag terms) outperformed the change-point model, shallow neural networks, and support vector machine regression (SVR). Our result implies that high resolution smart meter data should be used in place of monthly utility bills to more accurately evaluate retrofit savings. We further explored the relative contribution of the input variables to the random forest regression model using Shapley Value, a game theoretic variable importance metric. We found that the most important input feature is the energy consumption of the immediate past (or energy lag terms). We also found that solar radiation and weekend day indicators are more important than outdoor temperature. The improved model could provide better insights to portfolio managers in planning future energy retrofits. Policy makers could also use such models to evaluate the average energy saving potential for energy policy changes, such as the requirement of minimum insulation level, and lighting equipment efficiency