Abstract: The hydraulic support working resistance sequence in coal mine working faces during operations such as shifting, elevating, and conveyor pushing exhibits pronounced characteristics of strong non-stationarity, periodic disturbances, and complex multi-scale patterns, posing significant challenges for dynamic modeling and real-time pressure prediction. To address the modeling difficulties arising from long-term trend non-stationarity, short-term fluctuations, and abrupt changes in hydraulic support pressure time series, this study proposes a dual-path feature extraction and dynamic fusion prediction model. The model consists of a long-term trend modeling path based on an Informer encoder and a local path leveraging a multi-head differential attention-enhanced temporal convolutional network (MHDA-TCN). The Informer encoder introduces a probabilistic sparse attention mechanism to effectively reduce the computational complexity of self-attention and incorporates a self-attention distillation mechanism to capture long-sequence trend features. Meanwhile, the MHDA-TCN branch enhances sensitivity to high-frequency disturbances and abrupt structures via multi-head differential attention, combined with temporal convolutional layers to extract multi-scale local features. The features from both paths are adaptively integrated through a dynamic fusion module, balancing long-term trends with short-term perturbations to generate contextual representations for joint modeling. A parallel decoding strategy is adopted to predict multiple time steps simultaneously, significantly improving multi-step forecasting efficiency and long-sequence modeling capability. Experimental results demonstrate that, with a fixed window length of 128, the proposed model achieves optimal performance across multi-step prediction tasks. Specifically, for 20-step forecasting, the E_RMS and E_MA are reduced to 1.842 MPa and 1.362 MPa, while E_MAP and E_SMAP are kept below 8%, indicating robust long-term predictive accuracy. Comparative experiments further show that the proposed approach consistently outperforms baseline models such as Informer, Reformer, GRU-Informer, and XLSTM-Informer in short-, mid-, and long-term prediction tasks. In particular, for long-term forecasting, E_RMS, E_MAP, and E_SMAP decrease by 37.4%, 33.4%, and 32.3%, respectively, compared to the second-best model (XLSTM-Informer). These results demonstrate the proposed method’s strong real-time capability and robustness in modeling complex multi-scale characteristics of hydraulic support pressure sequences, providing technical support for intelligent roof pressure monitoring and early warning in coal mine working faces.