WPP-LLM: Multi-source heterogeneous data fusion for wind power prediction and early fault warning based on fine-tuned LLaMA-2

This paper proposes WPP-LLM, a fine-tuning framework built upon the open-source LLaMA-2 model, for wind power prediction (WPP) and predictive maintenance. WPP-LLM realizes unified processing for multiple data modalities by parameter-efficient fine-tuning (PEFT) of a pretrained large language model (LLM), without requiring full model retraining. First, a multimodal data alignment and tokenization strategy is designed to convert time series power data, text fault records, and visual defect features into a unified sequential representation. Subsequently, Low-Rank Adaptation (LoRA) is adopted for parameter-efficient fine-tuning of the pretrained LLaMA-2, so as to capture the complex spatio temporal and semantic correlations contained in heterogeneous inputs. Moreover, a time-series prediction head is designed for the WPP problem, while the operation and maintenance task is modeled as a multi-label classification task for anomaly detection and root cause analysis. Crucially, these early warnings directly trigger predictive maintenance workflows. Experimental results based on real data from a coastal wind farm show that WPP-LLM significantly outperforms traditional baselines and single-modal models in multiple metrics, which fully proves its superior ability in fusing heterogeneous data and improving the intelligent level of wind farm operation.