Measurement and Verification of Building Energy Savings
Assessing the effectiveness of building energy retrofits using measurement and verification (M&V) techniques is a critical requirement both for rate-payer funded energy efficiency programs and energy service performance contracts. Existing methods for M&V have several drawbacks. First, they do not capture nonlinearity, and hence cannot represent the nonlinear behavior of building performance. Second, they do not quantify uncertainty in calculated individual outcomes. Uncertainty quantification is crucial in evaluating energy savings with high confidence. And, importantly, they can be costly, requiring large amounts of baselining data, or alternatively, rigorous building modeling.
Argonne is working to develop robust M&V methods that use measured data and can be automated and integrated into building energy management systems. Leveraging our expertise in computational science, researchers are applying Gaussian Process (GP) models as a new statistical technique for M&V applications. GP models do not assume any relationships between independent and dependent variables, instead training a covariance matrix of input parameters that explains the energy outcomes. As a result, GP models can represent any trends of energy behavior suggested by the given data. Moreover, GP models account for uncertainty in the analysis process, and naturally quantify uncertainty associated with each prediction.
Early demonstrations have shown the power of the novel M&V technique to predict energy performance in an office building on-site at Argonne. The study demonstrated the capability of GP modeling to capture complex behavior, including nonlinearities, multivariable interactions, and time correlations. An associated simulation study showed that GP models require significantly fewer baselining points relative to linear regression for accurate energy predictions.
Building energy performance is non-linear. Using linear regression methods, piecewise models are needed to capture the behavior in different regimes, for example over the full range of outside temperatures. Using GP methods, the energy behavior in these different regimes can be predicted with one inference model. The uncertainty of the predictions can be reduced with higher time resolution data, for example by using hourly instead of daily data measurements.
This M&V method was developed in a project funded by U.S. Department of Energy's Building Technologies Program.