The energy consumption of buildings accounts for about one third of total energy consumption of our society, and the energy consumption of ice storage air conditioning system accounts for about half of energy consumption of buildings. Therefore, effective energy scheduling strategy of ice storage air conditioning system is of great significance to energy saving and energy cost reduction. In this paper, we propose a method to intelligently learn energy scheduling strategy of ice storage air conditioning system by using deep reinforcement learning technology. This method does not need to establish a building dynamics model, and directly uses temperature and other data for learning. Compared with the traditional rule-based and model-based methods, the efficiency is greatly improved.