The dynamic allocation of public resources faces challenges such as fluctuating demand, complex departmental coordination and frequent emergencies. Traditional static rules are difficult to balance resource utilization, response efficiency and regional fairness in a timely way. In this paper, a collaborative scheduling model combining digital twin, spatio-temporal supply and demand graph and deep reinforcement learning is constructed to transform the processes of community service, medical response, traffic connection, emergency supplies and administrative service windows into sequential decision-making problems. Method, the multi-scenario simulation is completed through the digital twin environment, and the graph neural network is used to encode regional demand, resource inventory, department load and event risk. A multi-objective reward function including task completion, resource utilization, fairness, response delay and scheduling cost is designed, and the policy update is realized by combining online feedback calibration. The experimental results show that the resource utilization rate of the proposed method reaches 89.7%, the task completion rate reaches 93.6%, the average response time is reduced to 18.6 min, the comprehensive performance is restored to 0.91 after the burst disturbance, and the task completion rate of cross-region migration remains above 92.0%. The research shows that this method can provide technical support for the fine allocation of public resources and the intelligent optimization of administrative efficiency.
