Buildings, which consume more than 70% of the electricity in the U.S., play significant roles in a smart-grid infrastructure. In order to automatically operate buildings to respond to signals from a smart-grid, high fidelity and computationally efficient building energy models are needed that can forecast a building's energy consumption under different operation conditions. Currently, most of the existing energy forecasting models are in three different categories, namely white box models (detailed physics based models), black box models (purely data driven models) and grey box models (hybrid models), among which black box models and grey box models are commonly used in on-line building control. However, typical data-driven black box models often require a long training period and are bounded to building operation conditions. On the other hand, creating even a simplified grey box model is often time consuming and need expert knowledge. An earlier study by Li and Wen (2014a) demonstrates the potential of using system identification methods to develop computationally efficient and accurate building energy forecasting models. However, there is a lack of systematic comparison in this early study to evaluate the proposed method against some other common building energy forecasting models. In this study, a set of model comparison criteria are proposed and used to evaluate the energy forecasting model generated from the reported system identification method and some other popular methods, including Resistance and Capacitance method, Support Vector Regression method, and Artificial Neural Networks method. A comprehensive EnergyPlus simulation model representing a mid-size commercial building is used to generate operating data for the model development and comparison in lieu of real building data. This study only focuses on cooling energy forecasting but the proposed method can be used for heating energy forecasting and whole building energy forecasting as well. Building operation data with and without sensor noise are generated for the comparison. In this comparison study, the model generated using the system identification method has the capability to achieve higher accuracy and extendibility under both of the noise-free and noisy conditions than those models generated by other methods.