Reduced-order modeling and simulated annealing optimization for efficient residential building utility bill calibration

Abstract

This simulation study applies the general framework described in BESTEST-EX for self-testing residential building energy model calibration methods. The National Renewable Energy Laboratory’s BEopt/DOE-2.2 is used to evaluate an automated regression metamodeling-based calibration approach in the context of monthly synthetic utility data for a 1960s-era existing home in a cooling-dominated climate. The home’s model inputs are assigned probability distributions representing uncertainty ranges, pseudo-random selections are made from the uncertainty ranges to define “explicit” input values, and synthetic utility billing data are generated using the explicit input values. A central composite design is used to develop response surface statistical models for the home’s predicted energy use. Applying a gradient-based simulated annealing optimization algorithm to the statistical “metamodels”, the calibration approach systematically adjusts values of the design variables and reduces disagreement between predicted energy use and synthetic utility billing data. Various retrofit measures are applied and used to assess accuracy of retrofit savings predictions resulting from using the calibration procedure. Substituting actual BEopt/DOE-2.2 model simulations with the statistical models reduces overall calibration procedure run-time while sacrificing only a limited degree of accuracy for retrofit savings predictions.