In order to improve the stability and accuracy of near-infrared spectroscopy (NIR) detection of total volatile basic nitrogen (TVB-N) in fresh mutton during storage (at 4 ℃, 8 ℃, 20 ℃), the selection of characteristic spectra and prediction models is the key step of NIR spectroscopy research. The 121 mutton samples were taken as experimental objects, the NIR spectra between 680 and 2 600 nm of fresh mutton samples were collected. The scattering correction methods, including multi scattering correction (MSC), standard normal transformation (SNV), and smoothing methods including Savitzky Golay convolution smoothing (SGS), moving average smoothing (MAS), and scaling methods including normalization, centring and auto scaling, were adopted to pretreat NIR spectra, and then PLS prediction models were built, by comparison, it is found that the spectra treated with SGS got the best modeling effect. Monte Carlo sampling (MCS) method and Mahalanobis distance method (MD) were used to eliminate 5 abnormal data of mutton spectra. The sample-set partitioning based on joint x-y distance (SPXY) algorithm was used to split 75% (87 samples) of the total samples as calibration set samples and the remaining 29 were validation set samples. The competitive adaptive reweighted sampling (CARS) algorithm, uninformative variable elimination (UVE) algorithm, improved uninformative variable elimination (IUVE) algorithm, successive projections algorithm (SPA) were employed to select characteristic wavelengths, and wavelength numbers were 14, 703, 144 and 15, respectively. The full spectra and the characteristic wavelengths selected by the four methods were taken as input variables to build prediction models, the results show that the performance of the model built with the wavelengths selected by CARS is better than the model built with the wavelengths selected by UVE, IUVE and SPA, and it shows that CARS method can effectively simplify the input variables and improve the performance of the prediction model. Compared with the UVE algorithm, the IUVE algorithm can select fewer wavelengths and improve the model’s performance. The PLS models, support vector machine (SVM) models and least squares support vector machine (LS-SVM) models were established with the selected characteristic wavelengths. The optimal prediction results of the calibration set are obtained by SVM models, in which the calibration determination coefficient ($R_{C}^{2}$) and root mean square error of calibration (RMSEC) of the CARS-SVM prediction model were 0.939 1 and 1.426 7, respectively. LS-SVM prediction model achieves the optimal prediction results of validation set, and the validation determination coefficient ($R_{V}^{2}$) and the root mean square error of validation (RMSEV) of IUVE-LS-SVM prediction model were 0.856 8 and 1.886 2, respectively. The simplified and optimized TVB-N prediction models for fresh mutton during the storage period are established based on NIR characteristic spectra, which provides reference and technical support for rapid and non-destructive detection of TVB-N concentration in fresh mutton.