An uncertainty aware transformer framework for wind power forecasting with multiscale attention and adaptive feature fusion

Abstract Wind power, as one of the most promising renewable energy sources, plays a pivotal role in the transition to a low-carbon energy system and the achievement of the "dual carbon" goals. However, wind power time series exhibit significant non-stationarity, strong randomness, and multi-scale temporal characteristics, which pose considerable challenges to the predictive accuracy and robustness of conventional forecasting models. To address these issues, this paper introduces the Uncertainty-Aware Forecasting Transformer (UAFformer) framework for wind power forecasting. UAFformer integrates several key mechanisms, including modal decomposition, multi-scale attention, dynamic weight reconstruction, and distribution modeling, to systematically enhance both the accuracy and stability of wind power predictions. Specifically, the model employs a hybrid decomposition strategy that combines Variational Mode Decomposition (VMD) and Singular Spectrum Analysis (SSA) to preprocess the original wind power sequence. This approach improves signal stationarity and extracts sub-modes with physical interpretations. Subsequently, a multi-scale temporal attention mechanism is introduced to capture dependencies across varying time scales. A dynamic weight learning module is designed to facilitate channel-adaptive reconstruction. In the feature fusion layer, an adaptive feature fusion gating structure is proposed to effectively integrate outputs from both the Transformer and BiGRU. At the output layer, an uncertainty perception module is developed to jointly model the mean and variance of the predicted power, thereby endowing the model with enhanced robustness and risk perception capabilities. Extensive experiments, ablation studies, and interpretability analyses were conducted using multiple real-world wind power datasets. The results demonstrate that UAFformer achieves competitive performance across multiple evaluation metrics, with notable improvements in accuracy and robustness. Specifically, UAFformer attains 8.7–12.3% lower mean absolute error during high-volatility intervals compared to state-of-the-art baselines, indicating enhanced stability under non-stationary conditions and strong potential for practical deployment.