Wind power prediction based on improved self-attention mechanism combined with Bi-directional Temporal Convolutional Network

Wind power output exhibits high randomness and volatility, making accurate forecasting challenging due to its inherent uncertainty. To address these challenges, this paper proposes a wind power prediction model incorporating Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), Sample Entropy (SE), Variational Mode Decomposition (VMD), and a Bidirectional Temporal Convolutional Neural Network-Transformer (BiTCN-TR). The model first applies CEEMDAN to decompose the wind power series into multiple modes. SE is then used to identify the most complex IMF, which is further decomposed using VMD to enhance stationarity in the wind power series. Finally, BiTCN-TR is applied to predict each component, and the predicted results are superimposed to obtain the final wind power forecast. To verify the reliability and applicability of the model, simulation experiments were conducted on cases from three different regions. The analysis show that for the low-volatility dataset, the MAPE is 1.662 % and the RMSE is 1.381 kW. For the medium-volatility dataset, the MAPE is 1.901 % and the RMSE is 0.58 kW. For the medium-to-high-volatility dataset, the MAPE is 1.870 % and the RMSE is 0.64 kW. These results demonstrate that the model maintains relatively high forecasting accuracy and stability across different volatility levels.