A multi-objective scheduling method for agile satellites based on nonlinear utility and deep reinforcement learning

The Agile Earth Observation Satellite Scheduling Problem is a multi-objective optimization task involving conflicting goals. These objectives include task priority, imaging quality, and energy consumption, which are highly correlated and often in conflict. Pareto optimisation is a key measure for evaluating the quality of complex multi objective solutions. However, conventional linear weighting methods neglect interactions among sub-objectives, generating suboptimal Pareto fronts with uneven trade-off distributions. This study proposes a multi-objective scheduling method for agile satellites based on Nonlinear Utility and Deep Reinforcement Learning. Within this framework, a Unified Scheduling Model (USM) was developed to provide a unified representation of constrained multi-objective observation tasks. Building on the USM, a contrastive learning–based nonlinear utility network was designed to capture inter-objective relationships via cosine similarity, thereby providing adaptive and diversified guidance for multi-objective trade-offs. Subsequently, a distributed training strategy combined with the Proximal Policy Optimization algorithm was employed to improve learning efficiency and policy stability under complex operational constraints. Results show that, compared with multi-objective deep reinforcement learning and multi-objective heuristic baseline algorithms, the proposed method achieved significant improvements in the Pareto volume and outperformed existing baseline methods in multiple performance indicators.