Abstract
High-fidelity building energy simulation models are software tools that are built on the well-established physical laws of thermal/hygric processes to model heating, cooling, lighting, ventilation, and energy use of buildings. They are highly nonlinear systems that involve a large number of subroutine calls and submodel switches during execution. To calibrate a building energy simulation model with good quality, parameter sensitivity analysis is well advocated, since it aims to identify those parameters in a specific building that hold more influence on the building thermal performance than others to facilitate shortening the lengthy cycle of model calibration procedures. Since simulation models are given in a large piece of source codes and encapsulate a series of submodels, the prevailing sensitivity analysis is mostly built within the framework of Monte Carlo simulation and statistics-based random sampling methods. It is computationally intensive. We propose to perform such analysis via a differential sensitivity analysis method that relies on the estimation of derivatives. A key technical challenge is that the high nonlinearity of the model prohibits any analytical differentiation, while numerical differentiation is too sensitive to step size and suffers from a truncation error. We, hence, propose to adopt an algorithmic differentiation method, which exploits the operator overload feature of object-oriented programs, to obtain accurate and robust numerical estimations of derivatives in an automated and computationally efficient way.
Note to Practitioners-This paper is motivated by the sensitivity analysis need arising from building energy simulation model calibration. Calibration procedures demand excessive parameter tuning and exorbitant simulations. A sensitivity analysis is thereby well advocated, since it helps identify those parameters that hold more effects to the building energy performance than others and should earn more attention when parameters are tuned. In this paper, we propose a local method via algorithmic differentiation for fast sensitivity analysis to facilitate building model calibration. In comparison with random-sampling-based methods, the proposed method can achieve similarly desired results, but is much more time-efficient.