Research on wind power prediction by integrating multi-scale decomposition and migration-enhanced deep learning architecture

With the rapid development of renewable energy, wind-power forecasting faces core challenges of strong volatility and non-stationarity, and traditional methods no longer meet accuracy requirements. To address this, a hybrid model, STL-HHO-BMGTL-OSRELM (SHBO), is proposed, integrating decomposition, transfer learning, and deep learning. Seasonal-trend decomposition (STL) splits the original sequence into trend, seasonal, and residual components to reduce non-stationarity. For the trend component, a Bi-Conv-MHSA-GRU (BMG) model is constructed. Bidirectional convolution (BiConv) extracts correlation patterns between adjacent time periods, and the multi-head self-attention gated recurrent unit (MHSA-GRU) enhances temporal-dependency modeling, mitigating the under-capture of long-term fluctuations. The Harris Hawks Optimization algorithm (HHO) performs dynamic parameter tuning to improve adaptability. Transfer learning transfers the trend-model weights to the seasonal-component prediction, leveraging data correlation to mitigate overfitting and improve generalization. Finally, the online sequence-regularized extreme learning machine (OSRELM) conducts real-time modeling of the residual component to correct short-term random fluctuations. Experimental results from a wind farm in Xinjiang, China, show that SHBO reduces root mean square error by 4.3% on average relative to the baseline across seasons, lowers mean absolute error by 4.7%, and decreases symmetric mean absolute percentage error by 22% on average, while improving stability to 0.98. These results indicate that the hybrid model effectively addresses the strong non-linearity and time-varying nature of wind-power forecasting through dimensionality reduction, multi-scale feature fusion, and dynamic residual correction, providing a scalable framework for forecasting highly volatile renewable-energy outputs.