Abstract: Ultrashort-term forecasting of wind power (WP) plays an important role in ensuring the safe and stable operation of power systems with high renewable energy ratio and promote WP consumption. Accurate forecasting results can promote WP consumption. In this paper, an ultrashort-term adaptive probabilistic forecasting of WP based on kernel density estimation (KDE) is proposed. The value of KDE with different band widths (BW) are generated according to different confidence levels, and the problem of poor robustness of the quantile obtained from the KDEs under different confidence levels is addressed. Ultrashort-term deterministic forecasting of WP based on improved bi-directional long short-term memory (BiLSTM) combines WP curves and data-driven in forecasting model. Thereby, the optimal BW KDEs are derived and an error-fitting model is constructed. This model can adaptively generate the optimal BW and construct forecasting intervals under different confidence levels. Finally, the proposed model is validated by the actual data, and the results show the superiority and effectiveness of the proposed model.