Memory-Enhanced Dynamic Self-Attention Communication Mechanism for Multi-UAV Base Stations Trajectory Planning

In large-scale temporary event scenarios, Uncrewed Aerial Vehicle base stations (UAV-BSs) provide flexible, on-demand wireless coverage to meet rapidly growing user demands. However, dynamic crowd movements and unpredictable user behaviors lead to highly uncertain environments. The absence of prior knowledge on user distribution, interference, and network load, coupled with UAVs' limited sensing capabilities, makes real-time trajectory planning especially challenging. Although reinforcement learning has shown potential in these scenarios, many existing approaches remain limited. They often make decisions based on instantaneous local observations or rely on unavailable centralized training information in a practical environment, limiting their adaptability and scalability. To address these limitations, we propose Mem-DSACM (Memory-Enhanced Dynamic Self-Attention Communication Mechanism), a lightweight reinforcement learning framework that integrates memory-assisted temporal reasoning with distance-constrained decentralized communication. In Mem-DSACM, the exchanged data, including relative distances and communication loads, is combined with historical observations and processed by a memory-based temporal reasoning module using a Transformer decoder. This enables context-aware and predictive decision-making under decentralized conditions. The simulation results demonstrate that the proposed method improves the system rate by 12.8%, the cumulative satisfaction 8.0% and the algorithm reduces the training time by 6.3% compared to the best baselines approaches, demonstrating superior performance in dynamic urban scenarios.