Wind Power Forecasting Based on Large Time Series Model

To address the insufficient forecasting accuracy in newly built wind farms due to a lack of historical data, we propose wind power forecasting methods based on the Moirai large time series model. For the few‑shot scenario, we fine‑tune Moirai on limited local samples, leveraging its generic time series representations learned from the large‑scale LOSTA dataset and adapting them to the target wind farm to mitigate overfitting under data scarcity. In the zero-shot scenario, we further pre‑train Moirai on a large‑scale, heterogeneous wind‑power dataset to build a large wind power forecasting foundation model with adaptive mapping dimensions. Experiments demonstrate that our few-shot approach significantly outperforms conventional deep neural networks under data–scarce conditions, meanwhile, our zero‑shot foundation model for wind‑power forecasting achieves 86.57% accuracy for ultra‑short‑term and 77.59% accuracy for short‑term forecasting, confirming its strong generalization and robustness across diverse wind farm environments.