STE-HOLNet：一种融合时空特征、动态概念漂移检测与自适应校正的风电功率预测新方法

Abstract Wind power is highly uncertain and nonlinear, and its time series often have multi-periodic features and concept drift, which poses a significant challenge for accurate prediction. In this paper, we propose a hybrid deep learning prediction model based on spatio-temporal feature enhancement with dynamic online correction (Spatio-temporal Enhanced Hybrid Online Learning Network, STE-HOLNet), which is tightly integrated with improved temporal coding and deep networks to achieve real-time, high-precision wind power prediction in real-time with high accuracy. First, an improved Time2Vec (E-Time2Vec) module is introduced to enhance the temporal feature representation; second, an improved TimesNetV2 network is designed to map one-dimensional time series into two-dimensional tensors and embedded with the Inception V2 module to extract features and identify changes in data distribution (concept drift) at multiple scales. The built-in concept drift detection mechanism detects abrupt changes in the statistical properties of the WT data in time to trigger subsequent adaptive training. Subsequently, the PS-HBPMamba adaptive training module was constructed by combining the Parameter Sharing Strategy (PS) with the Hedge Backpropagation online learning algorithm (HBP) and a novel Mamba network architecture. The module performs a local gradient update to the model when concept drift is detected. It dynamically adjusts the parameters to adapt to the new data distribution quickly, significantly improving the model’s stability and long-term prediction performance. Finally, the integrated error correction (EC) method corrects the prediction results to reduce the error accumulation phenomenon further and improve prediction accuracy. The experimental results show that the proposed model achieves excellent prediction performance on measured data from both inland wind farms in Xinjiang, China, and coastal wind farms in the United States. Compared with the current state-of-the-art benchmark models (PatchTST and inland wind farms, for example), the proposed model’s RMSE, MAE and SMAPE are reduced by 36.93%, 41.26% and 56.57%, respectively. The correlation coefficient, R2, is improved by 1.66%, significantly better than the existing state-of-the-art methods. These results verify the effectiveness of the combination of concept drift detection and adaptive online training and indicate that this study has achieved an important technical breakthrough in reducing the accumulation of prediction errors and enhancing the model’s ability to adapt to dynamic changes in data.