With the continuous development of renewable energy technologies and expansion of their system scales, a large number of renewable energy sources are being integrated into distribution networks. These renewable energy sources are increasingly managed in the framework of active distribution networks (ADNs). However, the expansion of the system scale and stochastic nature of user behaviors produce dynamic changes in the active and reactive power flows of these ADNs, which generate unpredictable operational stability problems such as voltage deviations. To address this challenge, a novel active and reactive power coordination control strategy is proposed for voltage control in ADNs. The strategy is based on a deep reinforcement learning framework, while introducing an attention mechanism to train an agent model for voltage control in an ADN in response to the system dynamics. To this end, this study first formulates an optimization problem for mapping between the active-reactive power and bus voltages in a system. Then, this problem is reformulated as a Markov decision process and solved using a local cross-channel interaction network-based soft actor-critic (LCCIN-SAC) algorithm. Simulation results on a modified IEEE 69-bus system demonstrate that the proposed strategy can successfully improve voltage deviation produced by rapid dynamic changes in a distribution network. The proposed strategy effectively ensures the stable and reliable operation of a distribution network with a high penetration of renewable energy sources.