Accurate ultra-short-term prediction of photovoltaic （PV） power is pivotal in mitigating the adverse effects of PV output uncertainty on power systems, thereby furnishing a dependable foundation for grid decision-making and scheduling. Therefore, a method for ultra-short-term prediction of photovoltaic output based on deep learning technology is proposed. Firstly, leveraging historical PV data, feature correlation analysis and K-means++ weather clustering are performed to discern pertinent patterns. And employing strategically positioned quantiles, a novel bidirectional long short term memory （BiLSTM） neural network is then constructed to capture bidirectional temporal dependencies. Then, a self-attention mechanism （SA） is introduced to dynamically emphasize key sequential information and a particle swarm optimization algorithm is integrated to optimize the parameters of the neural network. Optimal parameters are then assimilated into the BiLSTM-SA optimization framework for point prediction. Finally, through meticulous error analysis, a quantile regression （QR） is developed to delineate QR-BiLSTM-SA probability prediction model. The results show that the proposed method achieves accuracy exceeding 95% in ultra-short-term PV output probability prediction and has a good generalization ability, offering a robust foundation for the operation and scheduling of modern power systems.