When applying traditional attitude prediction methods in long-distance tunneling of shield, real-time update cannot be achieved due to changing geological conditions. To address this issue, an online shield attitude prediction model based on an incremental ensemble bidirectional long short-term memory network is developed. By learning new data in real time and dynamically updating model weights, the proposed method overcomes the limitations of traditional offline models in adapting to complex geological changes. Based on the shield boring data from rings 101 to 450 of the Hangzhou metro line 6, the historical data were filtered and weighted through the kernel mean matching algorithm, which were then combined with new data to form a dataset. Based on the previous-stage model, weak learners were generated through transfer learning, which were further combined into a strong learner using an ensemble algorithm. When the number of weak learners exceeded five, the weakest-performing learner was eliminated to ensure that the model maintained accuracy while controlling computational complexity, thereby meeting real-time updating requirements. The experimental results demonstrate that the proposed method considerably reduces attitude prediction error in segments with abrupt geological changes (rings 401－450), with the root mean square error reduced by 44.52% compared to traditional offline models, while training efficiency improved by 98.10%. Moreover, the model maintained high prediction accuracy across different stages and exhibited strong adaptability and robustness under complex geological conditions.