Abstract: With the continuous expansion of renewable energy integration scale, energy forms become more flexible and diverse, which presents new challenges to the dispatch operation of power systems. As the complexity and uncertainty of the system increase, the traditional optimization methods based on physical models are difficult to establish the accurate models for real-time and rapid solutions. In contrast, the deep reinforcement learning（DRL） can adaptively learn the scheduling strategies and make realtime decisions from historical experiences, avoiding the complex modeling process and coping with higher uncertainty and complexity in a data-driven manner. In this paper, the dispatch operation problems of new power systems are firstly introduced, then the principles and classification of DRL are described, and the advantages and disadvantages of various DRL methods to solve the dispatch decision problems of new power systems are analyzed. Finally, the trends that need further research are prospected.