Abstract: To facilitate the realization of automated shield tunnelling, an intelligent prediction model for shield attitude control parameters based on multi-scale convolution and attention mechanism is proposed. By learning from real-world shield tunnelling data from a tunnel in Zhuhai, this model constructed complex nonlinear mapping relationships among construction parameters, attitude parameters, and the pressures of six sets of thrust cylinders, thereby enabling intelligent prediction of the cylinder pressures required for shield attitude adjustment. On this basis, Bayesian optimization was employed to optimize the model hyperparameters. The effectiveness of the model architecture and the superiority of its prediction performance were validated through ablation experiments and model performance comparison. Furthermore, interpretability analysis method was used to reveal the influence of input features on the model prediction results. The research results show that the proposed model achieves excellent performance across all evaluation metrics on the test set: the maximum mean absolute error and root mean square error are only 1.0646 and 1.9822, respectively, and the coefficient of determination (R²) exceeds 0.98. The prediction performance outperforms multiple comparison models, demonstrating superior predictive capability. Concurrently, the ablation experiments confirm that each module of the model makes a significant contribution to the improvement of prediction performance.