Multiscale decomposition and attention-enhanced hybrid deep learning for wind power forecasting

Prediction methods that rely on a single temporal scale struggle to capture both the low-frequency trend information and high-frequency stochastic fluctuations in wind power time series, thereby limiting further improvements in forecasting accuracy. To improve the accuracy of wind power forecasting, this paper proposes a wind power prediction model based on multiscale decomposition and an attention-enhanced deep-learning approach. First, the original wind power sequence is decomposed into several intrinsic mode functions by complete ensemble empirical mode decomposition with adaptive noise. These intrinsic mode functions are classified by using fuzzy entropy, and intrinsic mode functions with similar complexity characteristics are clustered and reconstructed. Next, the high-frequency sequence is further decomposed using variational mode decomposition. Finally, the model uses temporal convolutional network combined with standard-deviation convolutional block attention module to extract local spatiotemporal features from decomposed sequences, followed by bidirectional long short-term memory network captures bidirectional temporal dependencies. The forecasting results of the decomposed sequences are aggregated to obtain the final prediction. Comparative experiments are conducted on wind power datasets from Turkey and France to evaluate the proposed model against other methods. The experimental results demonstrate that the proposed model achieves the best performance on both datasets in terms of error metrics, including mean absolute error, normalized mean absolute error, and root mean square error, while also attaining the highest coefficient of determination. These results indicate that, compared with the benchmark models, the proposed model can deliver more accurate wind power forecasts under complex wind speed fluctuation conditions.