Radial Basis Function Kolmogorov–Arnold Network for Coal Calorific Value Prediction Using Portable Near-Infrared Spectroscopy

The calorific value of coal is a key parameter for pricing, trade, and combustion management. Conventional bomb calorimetry provides accurate results but is time-consuming, labor-intensive, and destructive. Near-infrared (NIR) spectroscopy offers a rapid and non-destructive alternative, yet its application is limited by strong band correlations, nonlinear spectral responses, and the lack of interpretability in many predictive models. In this study, the Kolmogorov–Arnold Network (KAN) is applied to the prediction of coal calorific value, demonstrating its capability to describe nonlinear spectral relationships within an interpretable mathematical structure. Based on this framework, a Radial Basis Function KAN (RBF-KAN) is further developed by replacing the B-spline bases in the KAN with radial basis functions, allowing improved representation of localized and irregular spectral variations while maintaining model transparency. Using 671 coal-powder samples measured by a portable MicroNIR spectrometer, the RBF-KAN achieved an RMSE of 1.35 MJ/kg and an MAE of 0.92 MJ/kg under five-fold cross-validation, outperforming conventional regression models, deep neural networks, and other KAN variants. Analysis of RBF activations and spectral attribution maps indicates that the model consistently responds to characteristic O-H and C-H overtone regions, which correspond to known absorption features in coal. These results suggest that the RBF-KAN provides a practical and interpretable framework for on-site estimation of coal calorific value, complementing traditional calorimetric analysis.