The prediction of tunneling parameters for shield machines operating under complex geological conditions is significantly challenged by non-stationary data distributions, noise interference, and concept drift induced by abrupt stratum transitions. To address these issues, this paper proposes a residual test-time adaptive graph neural network (RTG-Net) for cutterhead torque prediction. First, to cope with the large magnitude discrepancies and rapidly varying statistical characteristics of torque signals across heterogeneous strata, a residual learning strategy is introduced to transform absolute-value prediction into incremental modeling. In addition, a robust optimization objective based on the Huber loss function is employed to alleviate the influence of anomalous noise and improve model stability. Second, a multi-scale dynamic spatiotemporal graph topology is constructed according to feature-space similarity. By coupling heterogeneous graph convolution with a cascaded bidirectional long short-term memory (BiLSTM) architecture, the proposed framework effectively decouples instantaneous geological-state representations from long-term temporal evolutionary dependencies. This design enables the extraction of correlations among discontinuous yet similar operating conditions, thereby enhancing feature representation capability under complex geological environments. Finally, to address concept drift caused by previously unseen strata, a test-time online feature adaptation strategy is developed. By introducing an experience replay mechanism, the input feature weights are dynamically updated, enabling adaptive model adjustment under unknown geological conditions without additional retraining. Engineering case studies demonstrate that the proposed method achieves a coefficient of determination (R2) of 0.992 7 and a mean absolute error (MAE) of 73.26 kN·m in single-step prediction tasks on cross-stratum test datasets, while reducing the root mean square error (RMSE) by more than 38% compared with baseline methods. Furthermore, RTG-Net maintains high accuracy and stability in multi-step forecasting tasks ranging from 2 to 5 step prediction horizons. The average inference time is only 3.13 ms per step, satisfying the real-time requirements of practical engineering applications. The proposed method effectively mitigates performance degradation in cross-stratum prediction and provides technical support for intelligent predictive-assisted driving of shield tunneling machines.