Abstract: With the integration of large-scale distributed generation into power systems, traditional reactive power optimization methods struggle to meet real-time control requirements in high-penetration renewable energy scenarios due to their strong dependency on models and low computational efficiency. To address this, a collaborative reactive power and voltage control method (GCN-MAAC) integrating graph convolutional network (GCN) and multi-agent attention critic (MAAC) is proposed. First, a distributed control architecture is constructed based on grid topology characteristics, and a reactive power-voltage optimization model considering renewable energy uncertainty is established, which is then transformed into a Partially Observable Markov Decision Process. Second, GCN is combined with a multi-head attention mechanism to design a multi-agent actor-critic algorithm framework. GCN captures the topological correlations of the power grid, while the attention mechanism dynamically weights critical node information to enhance learning efficiency. Additionally, the centralized training decentralized execution paradigm is adopted to address multi-agent cooperative decision-making. Finally, simulation experiments on the IEEE 118-bus system demonstrate that the proposed method outperforms traditional MADDPG and MAAC algorithms in both voltage deviation control and computational efficiency, providing an effective solution for reactive power and voltage control in power grids with high penetration renewable energy.