Exploration-guided adaptive learning from mixed-quality data for satellite multi-battery management

Mission-critical renewable energy power systems face severe challenges in data-scarce environments, as exemplified by satellite battery systems where harsh operating conditions including frequent eclipse-sunlight transitions, extreme temperature fluctuations, and space radiation exposure demand highly autonomous decision-making capabilities, yet high-quality operational data remains extremely scarce and prohibitively expensive to acquire. Conventional reductionist modeling approaches rely heavily on extensive labeled datasets, while existing machine learning methods exhibit fundamental deficiencies in multi-objective trade-offs, constraint satisfaction, and convergence guarantees, rendering them inadequate for zero-tolerance applications such as aerospace missions. This study proposes an exploration-guided adaptive learning paradigm that dynamically adjusts data utilization strategies based on quantified environmental uncertainty. The paradigm employs conservative exploitation mechanisms in regions with sparse state–action coverage, progressively enabling exploratory updates as distributional confidence accumulates, thereby achieving effective knowledge extraction from heterogeneous-quality datasets. Experimental results demonstrate that the proposed method achieves 3.6% improvement in power tracking accuracy and 2.5% enhancement in energy efficiency, with performance coefficient of variation reduced below 7.8%, validating the framework’s effectiveness and robustness for safety-critical systems in data-scarce scenarios. • Novel exploration-guided framework quantifies system understanding for adaptive learning. • Mixed regularization integrates behavioral cloning, exploration, and physics constraints. • Exploration-aware Actor-Critic dynamically balances conservatism and innovation. • Superior robustness across heterogeneous data quality in safety-critical systems.