CPLLM-WPF: A multi-scale prompting framework for generalizable wind power forecasting with LLMs

Wind power forecasting (WPF) is critical for promoting the integration of renewable energy into power systems. However, accurate and generalizable WPF remains challenging due to three core limitations in existing methods: 1) The inability to effectively incorporate textual information (e.g., task instructions, domain context) for enhanced accuracy and interpretability; 2) Difficulty in modeling complex multi-scale temporal dynamics in wind power data, including trends and sudden regime shifts; 3) Poor generalization performance across spatially heterogeneous wind farms, leading to high retraining costs and limited scalability. Large Language Models (LLMs) offer a promising solution due to their inherent textual knowledge, superior sequence modeling capabilities, and exceptional few-shot/zero-shot transfer potential. To harness these strengths for WPF, this paper proposes the CEEMDAN-guided Prompt-enhanced LLM for Wind Power Forecasting (CPLLM-WPF). We first use Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) to decompose each input feature into a set of Intrinsic Mode Functions (IMFs) and a residual, isolating trends and disturbances to provide the LLM with scale-specific inputs that mitigate volatility. A reprogramming layer then converts the numerical patches into token embeddings, and a task-aware prompt injects forecast instructions and component statistics, allowing the LLM to fuse textual knowledge with decomposed signals. Finally, Low-Rank Adaptation (LoRA) fine-tunes only the query and key projections, enabling adaptation to wind power domains with minimal computational overhead while preserving generalization. Extensive experiments on real-world wind farm datasets demonstrate that CPLLM-WPF consistently outperforms state-of-the-art baselines in both standard forecasting and zero-shot transfer scenarios, highlighting its accuracy, adaptability, and robustness.