Zero-shot forecasting of volatile wind power against data missing with large language model through attentive residual prompt tuning

Accurate wind power forecasting (WPF) is critical for maintaining supply-demand balance in smart grids. However, the inherent randomness of meteorological conditions and frequent operational data loss challenge robust trend prediction, especially for generalization to unseen wind farms. To address this, we propose an attentive residual prompt tuning approach, which constructs a multi-layer self-attention mechanism with shared attention parameters within the prompt tuning subspace, enabling LLMs to achieve zero-shot and robust WPF. A hard prompt generator is introduced to redefine WPF as a language modeling task, allowing the LLM to leverage its representation learning strengths for salient temporal feature extraction and anomaly detection. Meanwhile, a soft prompt adapter with attentive residual modules and bridge-shared parameters contextualizes the LLM for WPF tasks. The residual self-attention blocks model temporal dependencies both within and across wind farms, whereas the shared parameters facilitate feature transfer across temporal and spatial dimensions. This design mitigates reliance on complete input features, improving robustness to large-scale missing data. Furthermore, a hybrid soft-hard prompt fusion mechanism incorporates WPF-specific knowledge into the LLM, enhancing its zero-shot extrapolation capability under severe data loss. Extensive experiments on multi-region wind farms show our method surpasses seven state-of-the-art WPF approaches in robustness and accuracy under complex data loss conditions, and achieves superior zero-shot forecasting on unseen wind farms.