Gaussian process modeling for measurement and verification of building energy savings

• Published in: Energy and buildings Vol. 53; pp. 7 - 18
• Main Authors: Heo, Yeonsook, Zavala, Victor M.
• Format: Journal Article
• Language: English
• Published: Oxford Elsevier B.V 01.10.2012; Elsevier
• Subjects: Applied sciences; Building technical equipments; Buildings; Buildings. Public works; Computation methods. Tables. Charts; Energy management and energy conservation in building; Environmental engineering; Exact sciences and technology; Gaussian process modeling; Measurement and verification; Performance-based contracts; Retrofit analysis; Structural analysis. Stresses; Uncertainty; Measurement and verification; Uncertainty; Retrofit analysis; Performance-based contracts; Gaussian process modeling; Buildings; Verification; Modeling; Energy savings; Case study; Gaussian process; Strategy; Measurement method; Rehabilitation; Performance; Contract
• ISSN: 0378-7788
• DOI: 10.1016/j.enbuild.2012.06.024

► We present a Gaussian process modeling framework for measurement and verification. ► GP models can reliably determine energy savings and uncertainty levels. ► GP models can predict nonlinear and multi-resolution trends of energy behavior. ► Case studies demonstrated the applicability of the method for M&V practices. We present a Gaussian process (GP) modeling framework to determine energy savings and uncertainty levels in measurement and verification (M&V) practices. Existing M&V guidelines provide savings calculation procedures based on linear regression techniques that are limited in their predictive and uncertainty estimation capabilities. We demonstrate that, unlike linear regression, GP models can capture complex nonlinear and multivariable interactions as well as multiresolution trends of energy behavior. In addition, because GP models are developed under a Bayesian setting, they can capture different sources of uncertainty in a more systematic way. We demonstrate that these capabilities can ultimately lead to significantly less expensive M&V practices. We illustrate the developments using simulated and real data settings.