Abstract: The accurate prediction of tunnel boring machine (TBM) tunnelling parameters is of great importance for maintaining tunnelling process stability, mitigating construction risks, and enhancing operational efficiency. Traditional prediction methods, which rely on empirical expertise or static models, often struggle to handle the non-stationarity and long-range dependencies present in tunnelling data. A novel hybrid architecture—the FAN-TCN-Informer model—which integrates frequency-domain processing with multi-scale temporal modeling was proposed. The model innovatively combined Frequency Adaptive Normalization (FAN), Temporal Convolutional Network (TCN), and Informer. It leveraged FAN's frequency-domain decomposition capability to effectively mitigate distribution shift in non-stationary data, enhanced local feature extraction through TCN, and utilized Informer for efficient modeling of long-range dependencies, successfully achieving the deep fusion of multi-scale features. Using the actual TBM operational data from the +1095 m transportation tunnel at the Juxin Coal Mine in Guizhou Province as the research object, the effectiveness of the proposed method was validated by comparing the prediction performance of the model across different time steps. Experimental results demonstrate that the FAN-TCN-Informer model significantly outperforms commonly used forecasting models such as LSTM, GRU, and Transformer across multiple evaluation metrics. Particularly, in the 15-step prediction task, the mean absolute percentage error (MAPE) and mean absolute error (MAE) are only 4.4421% and 148.5953 respectively, demonstrating high prediction accuracy and stability.