To address the issue of insufficient accuracy in traditional prediction models caused by strong non-stationary and nonlinear characteristics of coal mine working face roof water inrush time series data, a dual optimization combination prediction model combining particle swarm optimization with successive variational modal decomposition and improved Newton-Rafson-Modified Transformer (PSO-SVMD-ENRBO-TRANSFORMER) is proposed. Using PSO algorithms to minimize average envelope entropy and adaptive optimization bandwidth constraint parameters of SVMD, the raw water inrush sequence is decomposed into multiple stationary intrinsic modal function components. An improved Newton-Rafsonian optimization (ENRBO) algorithm integrating Cubic chaotic inverse learning initialization, adaptive Newton-Rafsonian search rules, and Levey flight perturbations is proposed to perform global optimization of Transformer hyperparameters. Using the optimized Transformer, each component is independently predicted and reconstructed to obtain the final results. Using measured data from the 204 working face of a mine in the Binchang mining area, the results show that the model test set fits the goodness of R2 to 0.98, and the root mean square error (RMSE) is 8.59 m³·h-1, achieving an order of magnitude improvement from two-digit errors in traditional models to single-digit errors. Comparative experiments with traditional models such as BPNN, CNN, LSTM, and different decomposition methods, ablation experiments, and validation of adjacent working faces all show that this model has higher prediction accuracy and generalization capability, providing reference and guidance for predicting water inflow volume and water hazard prevention in mines under similar geological conditions.