In recent years, technologies such as artificial intelligence and distributed computing have significantly enhanced the operational performance of power distribution grids. However, significant challenges remain in areas such as computational task execution and data processing efficiency. An edge collaborative task offloading game optimization strategy based on deep reinforcement learning (DRL) was proposed to enhance task execution efficiency and overall computational performance for distribution networks. Firstly, a distribution network edge to edge collaborative task calculation model considering task priority was established. Furthermore, the task offloading problem was modeled as a Stackelberg game. Where the local server, as the leader, aims to optimize the comprehensive performance of task execution delay and computing cost. As a follower, the target uninstall server aims to provide computing services to maximize revenue. Then, the existence of Nash equilibrium in game model was proved. Finally, a game model solving method based on multi-agent deep deterministic policy gradient (MADDPG) was proposed. The case study results show that the local server utility value under the proposed method is increased by 66.7% and 52.4% respectively compared with the random offloading and greedy offloading algorithms. The average execution time of tasks is 14.4% less than that of random offloading. Furthermore, the load distribution of servers in the distribution network is more balanced, and the average load coefficient is reduced by 11.9% compared with the random unloading method.