Alum is an illegal additive that can improve the fragile characteristics of vermicelli. If the content of alum is excessive, it will directly affect the health of the body. This paper combines terahertz spectroscopy to explore a rapid detection method for alum in sweet potato starch. The spectral data of sweet potato starch, alum and their mixtures in the range of 0.5～7 THz were obtained by Terahertz Time Domain Spectroscopy （THz-TDS） at room temperature. Since the spectrum measured by 0～0.5 THz is noise, the absorption coefficient of the high-band region is large, and the signal-to-noise ratio is low, the absorption coefficient spectrum and the refractive index spectrum of the 0.5～2 THz band were selected for analysis. It was found that alum has obvious characteristic absorption peaks in terahertz band, which can be used as fingerprint features for material identification. Savitzky-Golay convolution smoothing, Baseline, Normalization were used for spectral pretreatment, and combined with partial least squares（PLS） a prediction model for alum content in sweet potato starch was established. The results showed that the principal component factors of the PLS model were 3, 3, 3, 2 using the original, SG smoothing, Baseline, Normalization spectral data, respectively. The correlation coefficient of calibration（rc） were 0.982, 0.980, 0.982, 0.984, respectively. The correlation coefficient of prediction （rp） were 0.982, 0.979, 0.982, and 0.987, respectively. The root mean square error of correction （RMSEC） were 0.011, 0.012, 0.012, and 0.011, respectively. The root mean square error of prediction (RMSEP) were 0.013, 0.014, 0.013, and 0.012, respectively. The PLS model had the best effect after normalization pretreatment. In order to compare and analyze the prediction accuracy of linear (PLS) and nonlinear (LS-SVM) quantitative model methods, the least square support vector machine was established using the spectral data of alum in the sweet potato starch after the same pretreatment method. For the prediction model, the radial basis function was chosen as the kernel function. The results showed that the LS-SVM model is the best after normalization preprocessing. The RMSEP of the prediction set was 0.004 7, and the correlation coefficient of the prediction set was 0.997 2. It was found that the LS-SVM prediction model for the alum content in sweet potato starch was more stable and more accurate. The content of alum in sweet potato starch was quantitatively analyzed by terahertz time domain spectroscopy combined with LS-SVM and PLS. The results showed that the LS-SVM with normalized pretreatment has better prediction effect than the PLS, which may be more nonlinear information in the mixture of sweet potato starch and alum. Studies have shown that terahertz time-domain spectroscopy combined with chemometric methods can provide a fast and accurate analytical method for the quantitative analysis of alum in sweet potato starch.