Abstract: To address the challenge of predicting deformation in tunnel surrounding rock within fault zones, this study aims to overcome the limitations of traditional methods in accurately capturing the nonlinear and non-stationary characteristics of surrounding rock deformation. The research employs an intelligent prediction model based on time convolutional networks, which extracts multi-scale temporal features through expansion convolution and residual connection mechanisms while integrating data from multiple monitoring sections. Validation was conducted using measured roof settlement data from eight monitoring sections along the Guangfo Ring Line GFHD-2 between days 11 and 20 after excavation. Results demonstrate that the average relative error across all sections ranged from 0.94% to 2.31%, with maximum errors below 3.5% and determination coefficients R² consistently exceeding 0.995. These findings confirm the model′s effectiveness in overcoming traditional method constraints and achieving high-precision temporal prediction. This approach provides a reliable and efficient intelligent deformation forecasting tool for tunneling through high-risk fault zones, offering significant engineering value for ensuring construction safety and guiding dynamic support design.