Mission Sequence Model and Deep Reinforcement Learning-Based Replanning Method for Multi-Satellite Observation.

Area of Science:

• Space Operations
• Artificial Intelligence
• Satellite Technology

Background:

• Increasing number of Earth Observation Satellites (EOSs) necessitates autonomous mission scheduling.
• Existing research often overlooks dynamic replanning for real-time sensor management in evolving environments.

Purpose of the Study:

• To address multi-satellite rapid mission replanning for dynamic batch-arrival observation requests.
• To maximize observation revenue while minimizing disruptions to original mission plans.

Main Methods:

• A framework integrating stochastic master-satellite mission allocation and single-satellite replanning using deep reinforcement learning.
• Utilizes mission sequence modeling with attention mechanisms and time-attitude-aware rotary positional encoding.
• Employs scalable embeddings for dynamic requests and a pointer network for efficient mission allocation.

Main Results:

• Achieved a 15.27% higher request insertion revenue rate and a 3.05% improvement in overall mission revenue rate.
• Maintained a 1.17% lower modification rate compared to state-of-the-art methods.
• Demonstrated faster computational speeds and significant performance improvements.

Conclusions:

• The proposed approach effectively handles dynamic observation requests for optimizing satellite sensor operations.
• The framework offers a robust solution for real-world satellite mission replanning challenges.
• Deep reinforcement learning and advanced modeling techniques enhance the efficiency and revenue of satellite missions.