Abstract: Mine water inrush has become one of the major problems restricting coal resource development in central and western China. To establish a water inflow prediction model suitable for mine working faces threatened by ultra-thick sandstone water disasters in western China, an integrated prediction model named GWO-iTransformer-SVM is proposed by combining the iTransformer model, an optimized variant of the mainstream Transformer model in current time series processing, with support vector machines (SVM), and using the grey wolf optimizer (GWO) to iteratively optimize parameters and weights during model training. The model first takes water inflow data of mine working faces as input, then conducts training on the iTransformer and SVM models respectively. The GWO algorithm is applied to iteratively optimize parameters and weights in the training and weighting processes of the two models, and finally outputs model evaluation indicators and water inflow prediction results. Compared with the GWO-iTransformer and iTransformer-SVM models, the generalization ability of the model is improved to a certain extent after iTransformer is weighted with SVM, and the prediction accuracy is also enhanced after parameter optimization via GWO. The training and prediction results of the GWO-iTransformer-SVM model show that the model achieves excellent performance in the training set, validation set and test set in terms of mean absolute error (E_MAE), mean absolute percentage error (E_MAPE), root mean square error (E_RMSE) and coefficient of determination (R²). Specifically, the E_MAE of the validation set and test set is 28.69 m³/h and 35.86 m³/h, the E_MAPE is 2.15% and 2.16%, the E_RMSE is 34.20 m³/h and 52.60 m³/h, and the R² is 0.95 and 0.93, respectively. Compared with other models, the proposed model features higher prediction accuracy and effectively avoids the complete non-fitting phenomenon observed in some models.