Coordinated attack missions involving multiple unmanned aerial vehicles (UAVs) require not only dynamic obstacle avoidance but also precise spatial-temporal synchronization to ensure simultaneous arrival and directional alignment at target locations. Existing planning methods often struggle to balance local adaptability with global coordination, particularly in environments that are cluttered, dynamic, or partially observable. To address these challenges, we propose a hierarchical framework that integrates deep reinforcement learning (DRL) with virtual guidance coordination for cooperative multi-UAV missions. In the first phase, each UAV independently navigates through complex environments by adjusting a disturbance-based fluid field, using an actor-critic DRL model to optimize path smoothness and obstacle avoidance. In the second phase, a virtual guidance point (VGP) mechanism enables decentralized yet coordinated convergence toward the target, ensuring synchronized arrival time and attack angle alignment across all UAVs. A shared reward structure is designed to guide both individual safety and group cooperation, enabling scalable and communication-efficient multi-agent learning. Extensive simulations in static and dynamic obstacle scenarios demonstrate that the proposed framework significantly outperforms baseline methods based on model predictive control (MPC) and artificial potential fields (APF) in terms of path feasibility, coordination accuracy, and mission success rate. The results highlight the effectiveness of combining fluid field modulation with reinforcement learning and VGP-based synchronization, offering a robust and extensible solution for real-time cooperative UAV operations in constrained environments.