Potato is the fourth important grain crop coordinated with wheat, rice and corn. At present, China is actively promoting the development of potato staple foods, but the uneven quality of potatoes has seriously hampered the process of the main food industry of the potatoes. Therefore, rapid non-destructive testing of potato quality is of great significance to the industrialization of processing. Domestic and foreign scholars have conducted a number of related researches on the detection of potato internal quality based on the visible/near-infrared diffuse reflectance principle. This method is commonly used, but the rough skin of the potato has a great impact on the detection. Another detection method is the transmission spectrum. This method can better reflect the internal quality information of the sample. However, the total transmission spectrum of the potato varies with the size of the sample and results in a large change in spectral intensity. Considering the above two reasons and average quality of potato, this study uses partial transmission spectrum as the detection method. This method can not only avoid the influence of the potato epidermis, but also obtain the internal information of the sample while maintaining the same path length. The spectral acquisition system consists of spectral acquisition units (spectroscopes and coupling lenses) and light source units (halogen lamps and lamp cups) which are arranged side by side. During testing, the two parts are attached to the sample surface to ensure that the spectral acquisition unit does not receive reflected light from the potato surface. Based on this system, partial transmission spectra of 120 potatoes are collected ranging from 650 to 1 100 nm. The prediction model of dry matter, starch and reducing sugar content was established using partial least squares regression after pretreat by detrend, multivariate scattering correction (MSC), standard normal variable transformation (SNV) and first-order derivative (FD). The result shows that the prediction models of dry matter and starch content using multiple scatter correction pretreatment are effective. The determination coefficients of validation set are 0.854 0 and 0.851 0, respectively, and the root mean square errors are 0.521 9% and 0.484 8%, respectively. The reducing sugar prediction model using first-order derivative pretreatment has the best result. The determination coefficients of validation set is 0.768 6 and the root mean square error is 0.025 1%. In order to optimize the model, three methods such as competitive adaptive reweighted sampling (CARS) are used to filter the characteristic wavelengths, and an optimized partial least-square prediction model is established. The result shows that the prediction effect of potato quality parameters has been greatly improved. The determination coefficient of validation sets for dry matter, starch, and reducing sugar prediction models after CARS screening are 0.877 6, 0.865 3 and 0.887 7, respectively. And the root mean square errors of the validation set are 0.449 2%, 0.930 2% and 0.016 7%, respectively. The use of CARS feature wavelength extraction can simplify the model and remove irrelevant variables and collinearity variables. This will improve the accuracy and stability of the model, especially for low-component content parameters such as reducing sugars. Finally, in order to verify the robust of the potato quality parameters prediction model, 30 potato samples are selected for external validation of the prediction model. The determination coefficients between model predicted values and standard physicochemical values of potato dry matter, starch, and reducing sugar are 0.849 9, 0.867 1, and 0.877 6, respectively. The root mean square errors are 0.660 9, 0.480 9, and 0.016 9, respectively. The average relative errors are 2.03%, 1.77% and 7.58%, respectively. The present study shows that the partial transmission spectrum carries the internal information of the potato and it is significantly related to the contents of dry matter, starch, and reducing sugar. The visible/near-infrared partial transmission detection system can achieve rapid and non-destructive prediction of multi-parameters of potatoes, especially good prediction results of dry matter content and starch content, but there is a large relative error in the prediction of individual samples with very low levels of reducing sugars. The next step of the study needs further optimization and improvement.