The mega-constellation satellite network consisting of thousands of satellites is a rising scenario for achieving global coverage. In the mega-constellation satellite network, it is important to ensure appropriate association between a satellite and its coverage area on the earth (denoted as a ‘cell’) to achieve the performance target, e.g., meeting the latency requirement. However, in the mega-constellation network, it is too complex problem to obtain the optimal association in real-time. In order to solve this problem, we propose an orchestration mechanism, called ODC, for dynamic management of the association between satellites and cells utilizing Deep Reinforcement Learning (DRL) model to adapt to time-varying set of visible satellites and their positions. Evaluation results show that ODC is able to reduce the communication latency by up to 15% compared with the conventional schemes, assuming the satellite constellations that are planned to commercialized.