Building heating, ventilation, and air conditioning (HVAC) systems consume large amounts of energy, and precise energy prediction is necessary for developing various energy-efficiency strategies. Energy prediction using data-driven models has received increasing attention in recent years. Typically, two types of driven models are used for building energy prediction: sequential and parallel predictive models. The latter uses the historical energy of the target building as training data to predict future energy consumption. However, for newly built buildings or buildings without historical data records, the energy can be estimated using the parallel model, which employs the energy data of similar buildings as training data. The second predictive model is seldom studied because the model input feature is difficult to identify and collect. Herein, we propose a novel key-variable-based parallel HVAC energy predictive model. This model has informative input features (including meteorological data, occupancy activity, and key variables representing building and system characteristics) and a simple architecture. A general key-variable screening toolkit which was more versatile and flexible than present parametric analysis tools was developed to facilitate the selection of key variables for the parallel HVAC energy predictive model. A case study is conducted to screen the key variables of hotel buildings in eastern China, based on which a parallel chiller energy predictive model is trained and tested. The average cross-test error measured in terms of the coefficient of variation of the root mean square error (CV-RMSE) and normalized mean bias error (NMBE) of the parallel chiller energy predictive model is approximately 16% and 8.3%, which is acceptable for energy prediction without using historical energy data of the target building.