As an essential measure to combat global warming, electric vehicles (EVs) have witnessed rapid growth. Meanwhile, thanks to the flexibility of EV charging, vehicle-to-grid (V2G) interaction has captured great attention. However, the direct control of individual EVs is challenging due to their small capacity, large number, and private information. Hence, it is the aggregator that interacts with the grid on behalf of EVs by characterizing their aggregate flexibility. In this paper, we focus on the aggregate EV power flexibility characterization problem. First, an offline model is built to obtain the lower and upper bounds of the aggregate power flexibility region. It ensures that any trajectory within the region is feasible. Then, considering that parameters such as real-time electricity prices and EV arrival/departure times are not known in advance, an online algorithm is developed based on Lyapunov optimization techniques. We prove that the charging time delays for EVs always meet the requirement even if they are not considered explicitly. Furthermore, real-time feedback is designed and integrated into the proposed online algorithm to better unlock EV power flexibility. Comprehensive performance comparisons are carried out to demonstrate the advantages of the proposed method.