Prompting large language model for multi-location multi-step zero-shot wind power forecasting

Wind power forecasting (WPF) is essential for the safe integration of renewable energy into the power grid. Despite the emergence of various predictors, they still face two challenges: 1) the predetermined forecasting horizon lacks the capabilities to make predictions over varying time steps, hindering flexible short- and long-term applications; 2) the chaotic nature of wind speeds under dynamic geographic and atmospheric conditions introduces mutation at certain time that complicates prediction and leads to large errors. To address these issues, a hard-soft hybrid prompt learning method is proposed to harness the potential of a large language model (LLM) for WPF. A hard prompt generator is proposed to redefine WPF as a language modeling task. It can leverage the powerful representation learning ability of LLM to capture inherent temporal features and reveal mutations within the WP data for accurate forecasts. A soft-prompt adapter with gated attention is proposed to align LLM with the WPF context via parameter efficient tuning. It can grasp the intrinsic spatial-temporal characteristics in and across wind farms with low computational complexity. Moreover, a hard-soft prompt fusion mechanism is designed to incorporate WPF-specific task information into the LLM, further triggering the zero-shot learning ability of LLM to achieve flexible prediction over varying forecasting horizons. Extensive experiments on wind farms distributed in multiple regions demonstrated that our method can flexibly perform both short- and longterm WPF tasks and robustly predict mutation events. It also outperforms the state-of-the-art deep predictors regarding generalization and accuracy.