Abstract: In order to improve the time-series modeling capability and multi-step prediction robustness of wind farm power prediction, a two-branch input ultra-short-term prediction method based on parallel bidirectional time convolution network (Bi-TCN) with hybrid signal decomposition and attention mechanism is proposed. Firstly, the data are divided into four seasons: spring, summer, fall and winter, and spring is selected as the experimental object because of the frequent alternation of cold and warm air in spring, which makes it more difficult to predict representatively. Subsequently, Pearson and Kendall correlation analyses were employed to screen meteorological factors with significant impacts on wind power variations, constructing a feature matrix of relevant variables. Following this, swarm decomposition (SWD) and improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) were collaboratively applied to process and construct a power feature matrix, achieving multi-resolution representation of the power sequence in the time-frequency domain. Finally, the power and associated variable characteristics are input to the Double-branches Bi-TCN-based PSHPA Attention and TCN-BiLSTM-SA architecture (DBT-BPPHA-TBS) model and the output is the last day of power prediction in the spring. The experimental results show that the DBT-BPPHA-TBS model is more stable than other wind power combination models in long time step prediction, with higher prediction accuracy and generalization ability.