In large-scale distribution networks, load transfer can effectively alleviate power supply shortages in individual partitions. However, due to the excessive network size, traditional mathematical optimization methods struggle to address this issue. First, a quantitative assessment model for partition transfer capability based on network simplification is proposed to select the scale of partitions for inter-sub-district operation. Second, an optimization method for inter-sub-district operation between distribution network partitions using deep reinforcement learning is introduced to quickly identify inter-sub-district operation support schemes. This approach improves the traditional reward function to enhance the agent's adaptability and generalization chracteristic across different load scenarios. Then, the dueling double deep Q-network(D3QN)algorithm is employed for policy learning to tackle the complexity of large-scale systems. Finally, simulations on the IEEE 33-node system and a practical 445-node system are conducted to validate the effectiveness of the proposed method.