Abstract: To enhance the robustness of multi-step wind farm power prediction, this paper proposes a dual-branch input ul-tra-short-term prediction method based on parallel bidirectional temporal convolutional network (Bi-TCN), hybrid signal decomposition, and attention mechanism. First, the dataset is divided by seasons. Spring is selected as the experimental sub-ject due to its frequent alternation of cold and warm air masses, higher prediction difficulty, and representative characteristics. Subsequently, Pearson and Kendall correlation analyses are adopted to screen meteorological factors that exert significant impacts on wind power variation, based on which a correlation variable feature matrix is constructed. Then, swarm decom-position and improved complete ensemble empirical mode decomposition with adaptive noise are utilized for collaborative processing to build the power feature matrix, realizing multi-resolution representation of the power series in the time-frequency domain. Finally, both the power feature matrix and the correlation variable feature matrix are fed into the proposed model (dual-branch temporal-Bi-TCN PSHPA hybrid attention-TCN-BiLSTM self-attention, DBT-BPPHA-TBS), generating the one-day power prediction results for the end of spring. Experimental results demon-strate that the DBT-BPPHA-TBS model outperforms other combined wind power prediction models in stability for long time-step prediction tasks, and achieves superior prediction accuracy and generalization capability.