With the rapid increase of in the number of electric vehicle (EV), battery swapping is becoming a promising idea because of its short service waiting time. However, in the face of the uncertainty of the power grid and EV behavior, it is difficult to achieve a forward-looking and fast-response scheduling in a large scale battery swap station (BSS). A new bi-level scheduling model is proposed to solve this problem, in which the upper level is built on a deep reinforcement learning (DRL) framework to optimally allocate power among the chargers, and the lower level is modeled as a series of MILP subproblems for dispatching power among the batteries in a charger. A prediction module is included in the DRL framework improve the foresight of the algorithm, and a safety module is designed to avoid unsafe actions. Experimental results indicate that the proposed approach has excellent performance in large scale problem solving. It reduces the operating costs of the BSS significantly while satisfying the maximum power demand constraint. This is able to provide more economic benefits for the BSS and help peak shaving and valley filling for the power grid.