Because of the high fluctuation and random nature of offshore wind power, time-series data contains not only global correlations but also a large amount of short-term local noise. In addition, existing probabilistic forecasting studies are commonly based on quantile regression methods which suffer from the quantile crossing problem and unreliable forecasting results. Therefore, this paper proposes an ultra-short-term offshore wind power probabilistic forecasting method based on a spectral attention and a crossing free joint quantile regression. First, combined with the long short-term memory (LSTM) model and spectral attention model, the global long-term dependence is mined and local noises are filtered in the time-series data. Then, a joint quantile regression loss function model is built to avoid the quantile crossing problem by dynamically adjusting the loss weights and sharing information among all the quantile regression tasks. Real case studies demonstrate that, compared to LSTM and TCN, the proposed model has improved the sharpness of quantile forecasting by over 6% and the continuously ranking probability score by over 10%, with superior prediction accuracy and the quantile crossing problem well addressed.