Abstract: To enable real-time optimization of cutterhead torque adjustment and prevent cutterhead clogging during the intelligent excavation of river-crossing shield tunneling, this study develops an Autoformer deep learning model for multi-step prediction of cutterhead torque in slurry shield tunnel boring machines, with hyperparameters optimized using Bayesian optimization (BO). First, the optimal input time step as well as the limit output step of the model were explored using historical time data of key shield machine parameters and cutter torque. On this basis, the model is analyzed by SHapley Additive exPlanations (SHAP) to enhance the interpretability of the model as well as the reliability of its application in real engineering. The study was carried out using actual data from the Wuhan river-crossing section, and the results show that: 1) The Autoformer algorithm provides support for accurate multi-step prediction of cutter torque, with an average R² of 0.946, an average RMSE of 189.149 kN.m, and an average MAE of 194.213 kN.m. 2) The Autoformer algorithm has stable prediction accuracy for a certain input time step, and the optimal input step for prediction is explored to be 16, and the limiting output step for prediction is 11. 3) Historical time data of the cutter torque is a key contributor to the future prediction results, and the cutter squeezing pressure also has a large impact on the prediction results.