Recurrent attention encoder–decoder network for multi-step interval wind power prediction

In the context of large-scale wind power grid integration, accurate wind power forecasting is crucial for optimizing grid scheduling and ensuring safe grid connection. This study proposes a recurrent attention encoder–decoder network for multi-step interval wind power forecasting, combining Numerical Weather Prediction (NWP) inputs with deep learning techniques. The approach leverages a sequence-to-sequence neural network and temporal attention mechanism, enabling better capture of latent patterns in historical data that are useful for future predictions, directly generating multi-step time series and final prediction intervals. Additionally, a moving window training scheme, integrating bifurcated sequences and hidden layers, is employed to organize historical data and improve the stability and performance of the sequence. Using offshore wind farm data, the wind speed and direction components (U, V) are decomposed, and experiments show that the proposed method outperforms existing methods in metrics such as a minimum PINAW of 0.119 and an average reduction of 19.37% in CWC. These results demonstrate high accuracy and reliability in interval forecasting, providing strong support for wind farm scheduling and grid optimization.