To improve the wind speed prediction accuracy of large-scale wind farm clusters and ensure the safe and stable operation of China’s power grid, a short-term wind speed hybrid prediction model for wind farm groups based on particle swarm optimization combined with projection pursuit clustering and NS-L-Transformer was proposed. Firstly, the collected wind speed dataset was processed by the methods of variational mode decomposition, depseudo-component removal and wavelet transform, and the wind speed dataset after filtering out noise interference was obtained. Secondly, considering the spatial correlation characteristics of wind speed among wind farm groups, according to the wind speed fluctuation characteristics, the particle swarm optimization based on projection pursuit clustering algorithm was used to analyze the spatial correlation between wind farm groups. Using the evaluation metrics obtained from the algorithm, an optimal classification of field group correlation was carried out based on their spatial correlations, and the classified high-dimensional wind speed dataset was constructed. Finally, the self-attention mechanism of the Transformer model combined with the gating unit mechanism of the LSTM model captured the local characteristics of the wind speed time series, and the NS-L-Transformer model was proposed to predict the wind speed of the constructed high-dimensional wind speed dataset with local characteristics. The wind speed data of a wind farm group in southeast China was selected for simulation analysis, and the results show that the prediction accuracy of wind speed prediction using the classified high-dimensional dataset is greatly improved compared with that of the single wind speed dataset. Furthermore, compared with the Transformer model, the NS-L-Transformer model exhibits reduced prediction errors, which validates the effectiveness of the hybrid prediction model proposed in this paper.