Abstract: Real-time prediction and control of shield machine attitude are crucial for ensuring construction safety. To address the issues of low prediction accuracy and ambiguous parameter selection in existing methods, this study proposes a hybrid prediction model based on TPE-XGBoost-GRU. The model considers four categories of factors influencing shield attitude, selects key features through the gain method in XGBoost algorithm, and optimizes the GRU model using Tree-structured Parzen Estimator (TPE) in Bayesian optimization. The superiority of the proposed model is verified by comparing the prediction performance of different hyperparameter optimization methods with deep learning algorithms. The results demonstrate that: (1) Historical shield attitude data play a critical role in attitude prediction; (2) During hyperparameter optimization, the number of hidden units and learning rate in GRU are key influencing factors, with importance weights of 0.36 and 0.30, respectively; (3) For prediction performance optimization, TPE outperforms random search and grid search, with maximum improvements of 41.1% in MAE and 12.0% in R2; (4) Under TPE optimization, the prediction performance of the three algorithm models ranks as GRU >LSTM > RNN.