Heating, ventilation, and air-conditioning (HVAC) system accounts for approximately 40% of total building energy consumption in the United States. Currently, most buildings still utilize constant air volume (CAV) systems with on/off control to meet the thermal loads of rooms. Such system, without any consideration of occupancy, may ventilate a room excessively and result in a waste of energy. Previous studies show that CO2-based demand-controlled ventilation are the most widely used strategies to determine the optimal level of supply air volume. However, many conventional CO2 mass balanced models assume equilibrium condition and may cause a time delay compared to the ground truth in determining the occupancy level. In this manuscript, a data-driven control strategy was developed to optimize the energy consumption of supply fans by using machine learning techniques to predict real-time occupancy as the active constraint. The experiment was taken in an auditorium located on a university campus. The HVAC system can be described as a single zone all-air system having a constant air flow. The result shows, after utilizing a new supply fan schedule, a maximum of 40% fan energy reduction can be achieved.