The granularity detection of pulverized coal in pneumatic conveying system is of great significance to the optimum control of coal mill. The traditional approach to detect the granularity of the coal granule is sampling the pulverized coal in the pipeline and applying sieving process. These steps are time-consuming and complex. Some rapid detection methods for the granularity of coal have already been released home and abroad, however, some drawbacks still exist, such as limiting the concentration of the pulverized coal to a low level during measurement and the instability of the testing equipment. Terahertz time-domain spectroscopy system (THz-TDS) is a newly developed non-destructive analytical technique. Compared with other spectrometry, THz-TDS has the superiority of low-energy photon, perm-selectivity and coherency. The previous research on the interaction between THz wave and granular medium indicated that the particle in granular medium has a strong influence on THz wave, which provides the technical feasibility of granularity detection of coal particle adopting THz-TDS. The propagation of the THz wave in granular medium could be regarded as a non-linear dynamic process involving complicated dynamic effect, leading to a THz signal combined with some certain chaos features. In this paper, the concept of non-linear chaos dynamic system was applied to the terahertz spectral analysis for the first time. Following this point of view, the detected THz signal was considered as a time series generated by a complex non-linear dynamic system and the interaction between THz wave and granular medium could be described by some chaos features. In the experiment, the coal was grounded and sieved into <38.5, 55~74, 74~88, 88~105 and 105~200 μm firstly. Then these pulverized coal samples were mixed with HDPE powder and compressed into tablets. The power spectral entropy, wavelet energy entropy, Box dimension, correlation dimension, skewness and kurtosis were extracted from the THz time domain signals of the six coal-HDPE tablets. Visually, the extracted chaos feature vectors showed a dependency with the granularity of the measured coal samples, so that the range of granularity could be roughly distinguished. However, the exact diameters of the measured coal samples remain unknown. Support vector machine (SVM) is a powerful tool for solving the small sample and non-linear classification problem. Appropriate parameters should be selected firstly, so that an accurate prediction model can be established by SVM. Particle swarm optimization (PSO) was used to optimize the parameters selection of SVM. The extracted chaos features were selected as inputs of the PSO-SVM to establish a regression model for predicting the grain size of the investigated granulated coal samples. The experimental result showed that the regression model trained by the chaos feature vectors had a worse performance of predicting the grain size of samples containing <38.5 and 38.5~55 μm coal granule than the model trained by the frequency depended extinction spectrum. This might be because of a relatively weak interaction between the THz wave and small grains that the chaos features of these samples are not significant. For the rest samples containing 55~74, 74~88, 88~105 and 105~200 μm coal grains, a better performance was achieved. Specifically, compared with the model trained by the extinction spectrum, for samples containing 74~88 and 105~200 μm coal grains, the prediction model trained by the chaos features obtained a lower RMSEC that declined by 29.48% and 26.14%, respectively and the RMSEP of this two samples declined by 88.62% and 56.86%, respectively. Overall, for the prediction model trained by the chaos features, the correlation coefficient between the predicted and actual particle diameter is 0.961 8, however, for the prediction model trained by the extinction spectrum, the correlation coefficient between the predicted and actual particle diameter is only 0.780 7. The RMSEP obtained by the model trained by the chaos features is 9.52, while the RMSEP obtained by the model trained by the extinction spectrum is up to 24.48. Furthermore, the elapsed time for modeling when adopting the chaos features declined by 43.19%. The research provides scientific basis and references for the application of granularitydetection of pulverized coal in pneumatic conveying system.