The maintenance scheduling (MS) of the distribution power network (DPN) needs to be based on the information from various units, and then discussions and coordination among multi-departments, such as maintenance and scheduling, need to be carried out. This process is dependent on human experience and the ability of departments to work together. First, the complex and coupled maintenance process is difficult to adapt to the growing demand for maintenance services in the DPN. Second, uncertainties increase the complexity of MS. How to achieve intelligent MS with fast calculation and analysis and high uncertainty decision-making is an urgent issue in the maintenance business. Therefore, this paper proposes a reinforcement learning scheduling method for MS. First, a multi-objective model for planning is established, taking into account the uniformity of maintenance work, the amount of adjustment to the planned time, and the economy cost. Then, the mathematical model was transformed into a Markov decision process, with the initial date of each plan as the action variable, and the planning agent was trained using deep Q network. Finally, the effectiveness of the method was verified using data from a real maintenance schedule. The results show that the method can adaptively solve the MS problem with uncertain durations, which expands the solution ideas for similar problems.