Abstract: Blasting operations constitute a critical production phase in open-pit coal mining, and the vibrations generated by these operations can pose significant risks to the surrounding environment and structures. Accurately forecasting the peak particle velocity (PPV) of blasting-induced vibrations is essential for mitigating damage and ensuring the safety of nearby facilities and personnel. This study employed Pearson correlation coefficients, Spearman rank correlation coefficients, and mutual information to analyze the key factors influencing PPV. By integrating the results from threshold screening and Lasso regression, the input feature values for the predictive model were determined. A deep learning framework combining Transformer and LSTM architectures was proposed to predict PPV in open-pit mining scenarios. This model uses the strengths of Transformers in handling time-series data and the memory capabilities of LSTMs, makes full use of its efficient modeling characteristics for time series data. Cross-modal splicing was utilized to fuse features, while residual connections and normalization technique were incorporated to improve model robustness. The findings indicate that blast center distance, elevation difference, blast area width, total charge, and row number are the primary determinants of PPV. Consequently, these five parameters were selected as input variables for the model. The performance of the Transformer-LSTM coupled model was evaluated using metrics such as the coefficient of determination (R² = 0.963), root mean square error (RMSE = 2.139), mean absolute error (MAE = 1.484), and mean absolute percentage error (MAPE = 32.431%). These metrics were compared with those of the reciprocal error method, XGBoost, IPSO-HKELM, and GA-LSSVM models. Specifically, R² increased by 0.63%, 7.96%, 7.24%, and 6.17% respectively; RMSE decreased by 51.19% and 143.53% relative to the reciprocal error method and GA-LSSVM; MAE decreased by 31.71%, 3.45%, and 56.26% compared to the reciprocal error method, IPSO-HKELM, and GA-LSSVM; and MAPE decreased by 37.63%, 45.94%, and 25.78% relative to XGBoost, IPSO-HKELM, and GA-LSSVM. Therefore, the Transformer-LSTM coupled model is superior to accuracy and stability in predicting PPV compared to the reciprocal error method, XGBoost, IPSO-HKELM, and GA-LSSVM models.