The HVAC systems in subway stations are energy consuming giants, each of which may consume over 10, 000 Kilowatts per day for cooling and ventilation. To save energy for the HVAC systems, it is critically important to firstly know the "load signatures" of the HVAC system, i.e., the quantity of heat imported from the outdoor environments and by the passengers respectively in different periods of a day, which will significantly benefit the design of control policies. In this paper, we present a novel sensing and learning approach to identify the load signature of the HVAC system in the subway stations. In particular, sensors and smart meters were deployed to monitor the indoor, outdoor temperatures, and the energy consumptions of the HVAC system in real-time. The number of passengers was counted by the ticket checking system. At the same time, the cooling supply provided by the HVAC system was inferred via the energy consumption logs of the HVAC system. Since the indoor temperature variations are driven by the difference of the loads and the cooling supply, linear regression model was proposed for the load signature, whose coefficients are derived via a proposed algorithm . We collected real sensing data and energy log data from HaiDianHuangZhuang Subway station, which is in line 4 of Beijing from the duration of July 2012 to Sept. 2012. The data was used to evaluate the coefficients of the regression model. The experiment results show typical variation signatures of the loads from the passengers and from the outdoor environments respectively, which provide important contexts for smart control policies.