Wind power generation, as one of the most critical clean energy sources and substitutes for fossil fuels, constitutes a pivotal component in achieving carbon neutrality. In power dispatch operations, the efficiency and accuracy of wind power forecasting serve as the most crucial and influential factors for grid management and operational decision-making. Current wind power prediction methodologies predominantly rely on historical single data source, while suffering from accumulating errors over extended forecasting horizons, leading to significant degradation in prediction accuracy that fails to meet grid scheduling requirements. To address the challenges of temporal dependency and error accumulation in wind power forecasting, we propose a deep learning framework based on multi-task dynamic learning and error correction. The model employs a ResNet-Transformer hybrid architecture, where Temporal Convolutional Networks (TCN) extract localized temporal features, while Transformer encoders capture long-term temporal dependencies through self-attention mechanisms, enhanced by residual connections for hierarchical feature reuse. Experimental validation on real-world wind farm datasets demonstrates that the calibrated predictions achieve a 45.686% RMSE reduction, evidencing the method's effectiveness under complex meteorological conditions. Comparative analyses further reveal substantial improvements in both short-term (<6 h) and extended-range (>24 h) forecasting performance compared to conventional baselines, establishing technical advantages for practical grid integration applications.