When representing building thermal characteristics for a large-scale district heating (DH) system, the traditional 1-order model is easy to solve but has a nonnegligible delay when conditions change. High-order models are widely used in building simulation, but their complexity and high time cost prevent them from being applied in DH simulation. The purpose of this paper is to find out an appropriate building thermal model for heating accident simulation. An unoccupied residential building located in northeast China was tested for 72 h heating outage and subsequent 32 h recovery. Based on the lumped parameter method and state space model, a high order was established by 4240 mass points to describe the tested building. By integrating mass points, the low-order models reduced the orders to eight, three, and one respectively. Frequency domain analysis revealed the similarities and differences among the models. The 1-order model showed a large deviation from the other three models. The subsequent time domain analysis of the 1-order model also simulated a root mean square error (RMSE) of 1.69, a delay of 6 h, and the lowest temperature drop during a heating outage. In both frequency and time domain analyses, the 8-order and 3-order models showed similar results approaching the high-order model, with an RMSE of approximately 0.50, a delay of 1 h, and small deviation. The simpler 3-order model could be used to estimate the indoor air temperature during a heating accident in large-scale DH systems. The further study on integrating the 3-order model and system identification method may overcome the shortcomings of model reduction.