A hybrid convolutional-LLM paradigm for robust wind power forecasting in data-scarce regimes

Accurate wind power forecasting is paramount for power grid stability, yet it faces a critical bottleneck in newly commissioned wind farms due to data scarcity, particularly the cold-start problem. Traditional data-driven models struggle to generalize in these low-resource regimes. To address this, we leverage the inherent generalization capabilities of Large Language Models (LLMs) to capture underlying physical dynamics from limited data. However, directly applying LLMs to continuous time-series remains challenging due to modality misalignment. In this paper, we propose a deep-learning enhanced LLM framework specifically engineered for wind power forecasting. Our approach introduces a lightweight convolutional adaptation head that aligns the semantic reasoning of a frozen LLM with the high-frequency patterns of numerical data. Quantitatively, our method demonstrates superior data efficiency compared to state-of-the-art baselines. Under extreme data scarcity with only 5% training data, our framework reduces the mean absolute error by approximately 27% compared to ablation baselines lacking the proposed prompt mechanism. Furthermore, unlike traditional models that suffer catastrophic performance degradation with limited data, our model maintains robust accuracy and achieves viable performance in zero-shot settings, offering a scalable forecasting solution for newly commissioned wind farms.