Due to the significant volatility and randomness of wind power data

low prediction accuracy is often observed with a single model in wind power prediction. To overcome this

an ultra-short-term wind power prediction method is introduced

based on modal decomposition and a combined neural network model. Firstly

the wind power data are processed based on the improved fully integrated empirical modal decomposition and sample entropy

which decomposes the unsteady series into smoother sub-sequences and reconstructs the high-frequency oscillatory component and low-frequency smooth component synchronously. Secondly

a hybrid prediction model for wind power based on an adaptive sparse self-attention mechanism is constructed. For the high-frequency oscillatory component with high complexity

the adaptive sparse Transformer model is used to fully explore the fluctuation information. For the low-frequency stationary components

the sequence features are fully extracted by the bidirectional gated recurrent unit model. Finally

the final prediction outcomes are derived by overlaying the forecast results of each component. Test was performed with actual data from a wind farm in Shandong

and the results show that

compared with other commonly used models

the proposed model's root mean square error and average absolute error has decreased by 2.644 MW and 2.42 MW

and the coefficient of determination has a notable 18.2% increase

implying it has a good prediction performance.