The complexity of the power system with high shares of renewable energy, driven by the uneven distribution of renewable resources and diverse regional energy demands, poses significant challenges for traditional planning methods. These approaches often struggle to capture the interregional dynamics and meet operational constraints needed for effective energy distribution. To address this, we propose the Constraint-Adaptive Graph-Convolution (CAG-C) Reinforcement Learning (RL) model. First, we decouple the large-scale planning problem using an RL-based framework that leverages the state transition process to break down interregional power flow optimization. This approach allows each region to be optimized independently. Graph Convolutional Networks (GCNs) are employed to extract region-specific data, where each node represents resource endowments and each edge represents power flow between regions, enabling the RL model to capture both regional and interregional dynamics. Finally, we apply a constraint-adaptive architecture that reduces the number of RL decision variables while ensuring all operational constraints are met. It works by adjusting each region's power input and output within feasible ranges. This ensures balanced power flows between regions while automatically satisfying requirements. Numerical results show that CAG-C can reduce planning costs and improve coordination across regions, offering a scalable solution for future power system planning.