Multi-step wind power prediction based on patch-wise attention mechanism

Effective feature representation is crucial for machine learning-based wind power prediction. Existing studies typically employ point-wise attention mechanisms to learn features at each time step of the wind power time series. However, this approach neglects the semantic correlations between consecutive time steps, limiting its ability to improve forecasting performance by learning rich semantic information. Therefore, a short-term multi-step forecasting strategy for wind power is proposed based on patch-wise attention mechanisms. By segmenting historical time series data into patches, patch feature vectors that can reflect semantic information such as time series trend are obtained. A Transformer encoder is constructed to achieve effective feature representation of complex semantic information. Based on the feature output by the encoder, a linear prediction output model is developed to achieve multi-step power forecasting. Case study results indicate that the wind power forecasting model based on the patch attention mechanism achieves a mean squared error (MSE) of 0.312 8 MW and a mean absolute error (MAE) of 0.384 8 MW, outperforming advanced models such as Transformer and Informer based on point-wise attention mechanisms.