Abstract
Wind power generation is a vital component of renewable energy, and achieving high-accuracy power forecasting is crucial for grid stability and sustainable operation. To address the highly nonlinear and complex temporal characteristics of wind power, this paper investigates a hybrid deep learning model, TCN-SENet-BiGRU-Global Attention. The model integrates a Temporal Convolutional Network (TCN), Squeeze-and-Excitation Network (SENet), Bidirectional Gated Recurrent Unit (BiGRU), and a Global Attention mechanism to construct a multi-level feature extraction architecture. Specifically, TCN efficiently captures both long-term and short-term temporal dependencies, SENet enhances the impact of key variables by adaptively adjusting channel-wise feature weights, BiGRU models bidirectional temporal context, and the Global Attention mechanism focuses on informative time steps to better track dynamic changes in wind power. Experiments on multiple real-world datasets from a wind farm demonstrate that the proposed TCN-SENet-BiGRU-Global Attention model achieves consistently lower prediction errors and more stable performance than several representative baseline models, indicating its good robustness and promising application potential for complex short-term wind power forecasting tasks.