Abstract: Aiming at the problems of poor stability, long detection cycles, and difficulty in achieving rapid on-site identification in traditional coal classification methods, a coal classification method based on near-infrared (NIR) spectroscopy and an adaptively optimized broad learning system (AO-BLS) is proposed. A preprocessing approach combining Savitzky-Golay (SG) filtering with standard normal variate (SNV) transformation is employed to correct random noise, scattering interference, and baseline drift in the raw spectra. Subsequently, a broad learning system (BLS) is constructed, and the particle swarm optimization (PSO) algorithm is introduced to optimize the key hyperparameters of the model. By integrating dynamic inertia weights, dynamic learning factors, and an elite perturbation mechanism, the AO-BLS method is proposed. Taking four typical coal types—anthracite, coking coal, other bituminous coal, and lignite—as the research objects, a coal classification study along with comparative experiments is conducted using 1450 sets of NIR spectral data. Classification Accuracy, Precision, Recall, and F1-score are selected as evaluation metrics to comprehensively measure the model's performance. The results demonstrate that the proposed AO-BLS method achieves a classification accuracy of 96.55% on the test set, outperforming comparative models such as K-nearest neighbors (KNN), support vector machines (SVM), and the conventional BLS in overall performance. It can accurately identify anthracite and coking coal, with only a small number of sample misclassifications occurring between other bituminous coal and lignite, which possess similar coal ranks. The combination of NIR spectroscopy and the adaptively optimized broad learning system can provide reliable technical support for the rapid and non-destructive classification of coal types, while enriching research on the high-precision and stable classification of coal types with similar categorical features.