[Objective] Although mainstream deep learning models have achieved excellent performance in wind power forecasting，they rely heavily on large amounts of historical data. In small-sample scenarios，such as newly built wind farms or situations involving privacy protection where data availability is limited，their prediction performance often declines significantly. To address this issue，a small-sample wind power forecasting method based on a time-series large model is proposed. [Methods] First，mean scaling and uniform binning techniques are employed to perform per-sample normalization and quantization of wind power time series，converting them into discrete token sequences so that they match the input format required by large models. Second，a text-to-text transfer transformer architecture is introduced as the core forecasting model. By leveraging the general representation capability obtained through pre-training on large-scale and multi-domain time-series data，the model is able to extract deep temporal features from wind power sequences. Finally，for the small-sample data from a specific wind farm，a lightweight fine-tuning strategy based on low-rank adaptation is adopted to efficiently adapt the model while preserving the knowledge learned during pre-training. [Results] Simulation results demonstrate that under data-scarce conditions，the proposed method outperforms several mainstream deep learning models in wind power forecasting tasks across different seasons，forecasting horizons，and regions. Compared with benchmark models，the root mean square error and mean absolute error are reduced by 15.18%-33.59% and 10.18%-34.94%，respectively. [Conclusions] The proposed method effectively uses the transfer learning capability of time-series large models to achieve high-accuracy wind power forecasting under small-sample conditions. It provides a feasible and efficient technical solution for high-precision wind power forecasting in data-limited scenarios such as newly built wind farms.