Uncertainty-Aware Domain-Specific Large Language Model for Maintenance Decision-Making on Wind Farms

Abstract

Wind farm maintenance planning requires better strategies that reduce operation and maintenance cost under uncertainty. Traditional methods can produce feasible solutions, but they usually rely on explicit modeling and scenario-specific solving setups, which limits flexibility when scenarios or uncertainty settings change. General large language models can understand task descriptions, but they often generate plans that cannot be executed. To address this problem, we propose Unc-LLM, a domain-specific large language model for uncertaintyaware maintenance decision making. Unc-LLM is trained via supervised fine-tuning on labeled maintenance optimization instances and learns to generate maintenance strategies for wind farms with multiple fault modes, heterogeneous repair costs, and uncertain conditions. Compared with the untuned Qwen3.5-9B baseline, Unc-LLM generates executable plans much more reliably while maintaining good decision quality in complex scenarios. Experimental results show that Unc-LLM achieves high executability on fixed-scenario benchmarks, with mean absolute percentage error (MAPE) below 2% among implementable plans. It also shows meaningful cross-scenario transfer, and the model trained on the more complex setting shows the strongest transfer performance, although the hardest structural shifts remain challenging. In heuristic reference comparisons on the two fixed-scenario benchmarks, Unc-LLM records MAPE values of 0.9408% and 1.8872%. This further demonstrates the strong performance of Unc-LLM relative to most heuristic baselines.