As mining activities continue to intensify, underground roadways encounter unprecedented stability challenges due to increasingly complex stress environments, dynamic geological variations, and anthropogenic disturbances. Accurate prediction of roadway deformation is essential for ensuring mining safety and optimizing operational layouts. However, conventional single neural network models often struggle to effectively capture both abrupt local features and long-term evolutionary trends in nonlinear displacement time series. To address these limitations, this study introduces an innovative hybrid model combining a Temporal Convolutional Network and Long Short-Term Memory(TCN-LSTM-AddAttn) architecture, enhanced with an additive attention mechanism, to achieve high-precision predictions of roadway deformation. The proposed architecture employs a parallel framework to leverage the strengths of both TCN and LSTM. An additive attention mechanism is incorporated to dynamically prioritize critical patterns from both networks. The model employs learnable parameters to calculate feature similarity and utilizes the Softmax function to generate normalized weights, facilitating the adaptive fusion of multi-scale representations. Validation of the model is conducted using displacement data from four monitoring points(W1430-10, W1430-11, W1480-7, W1530-11) across various roadways in the Yunnan Zizou iron mine. Data preprocessing involves the removal of outliers using a Hampel filter, interpolation of missing values via cubic spline, and min-max normalization to standardize input scales. The processed data are divided into training, validation, and test sets in an 8∶1∶1 ratio. Hyperparameters, including TCN channels (32), LSTM hidden dimensions (64), batch size (32), and learning rate (0.001), are optimized through grid search to ensuring generalization across diverse mining scenarios. Experimental results indicate that the TCN-LSTM-AddAttn model outperforms standalone TCN, LSTM, and the TCN-LSTM hybrid models. In the case of W1430-10, the TCN-LSTM-AddAttn model demonstrates a Mean Absolute Error (MAE) of 0.0292 mm, representing a 47.95% reduction compared to the TCN-LSTM model. Additionally, it achieves a Root Mean Square Error (RMSE) of 0.0396 mm, marking a 37.34% reduction, and a Symmetric Mean Absolute Percentage Error (SMAPE) of 0.0337, indicating a 47.91% reduction. The Adjusted R² (R adj) value of 0.9861 suggests near-perfect prediction accuracy. For W1430-11, the model records an MAE of 0.0282 mm (28.79% reduction), an RMSE of 0.0366 mm (27.52% reduction), a SMAPE of 0.0738 (28.70% reduction), and an R adj of 0.9799. Comparable improvements are noted for W1480-7 and W1530-11, with prediction errors consistently remaining below 0.1 mm. By incorporating multi-scale feature decoupling and dynamic weighting, the proposed model offers robust technical support for assessing mine roadway stability, identifying risk zones, and providing early safety warnings.