Stratospheric airship trajectory planning via temporal perception and dual-source learning

Stratospheric airships are lighter-than-air platforms offering long endurance for missions like communication relay and monitoring, but their performance highly depends on trajectory planning methods. Existing trajectory planning methods, however, often struggle to exploit the complex spatiotemporal patterns in weather forecasts or address the critical data misalignment between predictions and reality. This raises the fundamental question of how to design an autonomous agent that can proactively plan trajectories by leveraging imperfect future weather information. To solve this, we propose a novel deep reinforcement learning framework featuring a Temporal Perception Module (TPM), which uses a Transformer-based look-ahead mechanism to allow a Soft Actor-Critic (SAC) agent to anticipate weather changes. We also introduce a dual-source learning paradigm, where the agent observes Pangu-Weather forecasts while interacting with an environment simulated from higher-fidelity ERA5 reanalysis data to improve robustness. In comprehensive experiments, the TPM-SAC framework achieved an average mission success rate of 74.64%, representing a significant 11.14% absolute improvement over the baseline SAC model. Furthermore, we identified an optimal look-ahead horizon of 4 hours, which best balances predictive benefit and accumulated forecast uncertainty. This research contributes a more robust and efficient trajectory planning approach, thereby enhancing mission endurance and the operational reliability of stratospheric airships in dynamic, real-world environments.