Traditional coal quality analysis methods are often characterized by cumbersome procedures, long testing cycles, and high resource consumption, making them inadequate for the demands of modern energy systems in terms of efficient utilization and safety assurance. Portable near-infrared (NIR) spectroscopy, with advantages such as rapid, non-destructive, and on-line detection, has broad application prospects in coal quality analysis. However, issues such as spectral peak overlap, high inter-variable correlation, and sensitivity to sampling conditions still limit modeling accuracy and stability. This paper proposes a multi-task feature decoupling U-shaped network based on portable NIR spectroscopy for the collaborative prediction of coal volatile matter and calorific value. First, an iterative outlier elimination strategy is developed using the Rairda criterion and Euclidean distance to enhance the reliability of modeling data. Second, various spectral preprocessing methods are compared, verifying the superiority of standard normal variate transformation in suppressing baseline drift and scattering effects. Next, a Unet-based shared-parameter module with an encoder–decoder architecture and skip connections is employed to efficiently extract shared features. Finally, a multi-task feature decoupling module is introduced, where orthogonal constraints and auxiliary prediction heads jointly optimize the separation of task-specific features while maintaining both inter-index correlation and specificity. Experimental results on 600 coal samples show that the proposed model achieves a root mean square error (RMSE) of 1.4255, mean absolute error (MAE) of 0.9600, and correlation coefficient (R) of 0.8086 for volatile matter prediction, and an RMSE of 1.2954, MAE of 0.9320, and R of 0.8584 for calorific value prediction—significantly outperforming various traditional machine learning and deep learning models. Furthermore, noise interference experiments confirm the model’s robustness and generalization capability under complex sampling conditions. This study provides an efficient, accurate, and highly adaptable technical approach for portable, multi-index, on-line coal quality detection.