Abstract: With the rapid development of dual-carbon targets, many distributed power sources, represented by wind power and photovoltaics, are being connected to distribution networks. This will further exacerbate the intermittency and volatility of power output. Dynamic reconfiguration of active distribution networks constitutes a complex, high-dimensional, mixed-integer, nonlinear, and stochastic optimization problem. Traditional algorithms exhibit numerous shortcomings in addressing this issue. By integrating the advantages of both deep learning and reinforcement learning, the deep reinforcement learning algorithm is highly suitable for formulating dynamically reconfigurable strategies for active distribution networks, which are currently of great concern. This paper first summarizes the characteristics of the active distribution network of the new generation power system, and analyzes the progress and challenges of the current research on the dynamic reconfiguration of the active distribution network in mathematical models. Secondly, the coding method of the distribution network dynamic reconfiguration is discussed, and the deep reinforcement learning algorithm is systematically reviewed. Furthermore, the shortcomings of the existing algorithms in dealing with the dynamic reconfiguration of the active distribution network are analyzed, and the research status and advantages of the deep reinforcement learning algorithm in the dynamic reconfiguration of the active distribution network are summarized. Finally, the future research directions for the dynamic reconfiguration of active distribution networks are presented.