Aiming at the problem that traditional TCN models in industrial process dynamic modeling tend to ignore the continuity of time series and local dependencies, a dynamic soft-sensing model called CAFF-GTCN is proposed. A novel cascaded attentional feature fusion module is designed to improve the residual connections, which utilizes self-attention mechanisms and multi-scale channel attention mechanisms to fuse features extracted from different receptive fields, avoiding the loss of important information. Meanwhile, a gating mechanism is introduced to modify the dilated causal convolution, combined with the SELU function, to enhance the ability to extract dynamic features. Experimental results show that the proposed method significantly improves prediction accuracy: in the penicillin fermentation simulation experiment, compared to the traditional TCN model, the CAFF-GTCN model reduce RMSE and MAE by 45.1% and 49.4%, respectively, while increasing R2 from 0.992 3 to 0.998 9. In the sulfur recovery process experiment, the CAFF-GTCN model reduce RMSE and MAE by 38.2% and 42.7%, respectively, while increasing R2 from 0.750 3 to 0.846 4. The experimental results verify the effectiveness and superiority of the proposed method in dynamic feature extraction and prediction accuracy.