Abstract: To ensure the safe and efficient extraction of coal mines and reduce sudden roof accidents, this study proposes a novel multivariate long-term sequence mine pressure prediction model—the Empirical Mode Decomposition Linear Model (EMD–Mixer). Unlike most fixed-length single-feature mine pressure prediction models, this model first introduces the Empirical Mode Decomposition (EMD) method to separate periodic and trend components from the mine pressure signals. It then combines these with the Long-Term Forecasting Linear Layer (LTSF–Linear) to form a module for extracting temporal features. Additionally, a time and channel mixing strategy is designed, utilizing a channel feature module to handle multivariate mine pressure data and process nonlinear relationships. The final prediction results are obtained by adding the residuals of the time and channel module outputs to the input data. In the experiments, the historical window was set to 36 time units, and tests were conducted for prediction lengths of 24, 36, 48, and 60 time units. The results indicate that the EMD–Mixer model exhibits excellent performance and stability across short to medium-term prediction ranges. The model was compared with the LTSF–Linear, MTS–Mixer, and the commonly used Long Short-Term Memory (LSTM) model in mine pressure prediction. Four evaluation metrics were used to assess the prediction results: Mean Absolute Error (E_MAE), Mean Squared Error (E_MSE), Symmetric Mean Absolute Percentage Error (E_sMAPE), and R-squared (R²). The results show that the EMD–Mixer model demonstrates higher prediction accuracy and stability across all metrics. The EMD–Mixer model is simple in structure, has strong generalization capabilities, and can more effectively adapt to multivariate mine pressure data of different levels in various scenarios. This provides an important research approach for the safe and efficient production of coal mines and the early warning of roof accidents.