High-fidelity building energy simulation models are established on well-known physical laws and attempt to faithfully represent a real building. The numerical solution of differential algebraic equations that interpret thermal balance of buildings is included in building energy simulation models. It thereby leads to highly discontinuous systems that involve a large number of sub-routine calls and model switches during the execution. To acquire a building energy simulation model with good quality, parameter sensitivity analysis is well-advocated since it aims to target those parameters from the parameter pool in a specific building that hold more influence on the building thermal performance than others. Since building energy simulation models are given in a large piece of program codes and encapsulate a series of sub-models, the existing sensitivity analysis is built on Monte Carlo simulation and statistics-based random sampling methods only, e.g., Monte Carlo sampling and Latin Hypercube sampling methods, which are extremely time-consuming. We propose to perform the sensitivity analysis of a first-principle high-fidelity building energy simulation model via a straightforward differential sensitivity analysis method that relies on the estimation of derivatives. A key technical challenge is that the complexity of the model prohibits the analytical differentiation, while the numerical differentiation is sensitive to step size and suffers from the truncation error. We, hence, propose to adopt an automatic differentiation method, which exploits the operator overload feature of object oriented programming language, to obtain accurate numerical estimations of derivatives in an automated and computationally efficient way.