Abstract: Excessive shield posture deviations can severely compromise the safety of shield construction and project quality. To accurately predict shield posture deviations and provide decision support for early correction, a machine learning algorithm combining Empirical Mode Decomposition (EMD) with fixed component number and Gate Recurrent Unit (GRU) is proposed for grouped prediction of shield posture deviations in different directions. The results show that EMD can decompose irregular shield posture deviation signals into several regular components, and GRU can extract temporal dependencies in the input data, enabling precise prediction of shield posture deviations in different directions. Using a fixed number of EMD decomposition components and grouped prediction method ensures the independence of the prediction and training sets. After comparison with other models, the EMD-GRU model proved to be the optimal model. The analysis of the effects of different input parameter combinations on prediction results verifies the interdependence of posture deviations in the same direction (horizontal or vertical).