With the advancement of energy transformation and the construction of new-type power system, the importance of operation control and optimization technologies for distribution networks is gradually highlighted. Mechanism-driven methods are insufficient to fully solve problems such as inaccurate parameters for modelling, uncertainty in power sources and loads, and coordination of large-scale control resources. As a result, data-driven technologies represented by deep reinforcement learning have become a research hotspot. This paper summarizes the problems of operation control and optimization in distribution networks, and then reviews the current application routes of deep reinforcement learning, especially multi-agent deep reinforcement learning, in three typical scenarios: Volt/Var control, distribution network reconfiguration and restoration. Finally, two key breakthrough directions are proposed: technologies for improving the convergence efficiency and enhancing decision trustworthiness in distribution networks.