Misusing the wrong grades polymer during the polymer processing in the same production line may lead to poorer product performance and a lower qualification ratio. The traditional methods identifying the grades from same kind of polymer are usually time consuming and hysteretic. There has not yet been discovered a fast and real time method for grade identification. In this work, 5 different grades of GPPS were the research object. An in-line near-infrared spectral measurement system installed on the extruder was developed. Near-infrared spectroscopy was combined with chemometrics and machine learning algorithms. The different grades of GPPS could be fast and in-line identified during the extrusion process. First, the in- line near-infrared spectra of GPPS melts of 5 different grades were collected in real time by the developed system with a spectral range of 900~1 700 nm. After spectrum analysis, a K-means clustering algorithm in combination with PCA was performed to verify the distinguishability of in-line near-infrared spectra for different grades. Last, PLS-DA and RF algorithm were used to establish the grade identification models respectively, and the identification ability of these two models was compared. The results show that: ①After baseline correction, maximum and minimum normalization, and 7-point moving average smoothing, the characteristic peak values at 1 207, 1 388, 1 407, 1 429 nm of the in-line near-infrared spectra change in a step-like manner with the change of grades. With the first three principal components scores as input variables, the clustering accuracy by K-means can reach 88%. It shows the distinguishability of the in-line near-infrared spectral data of different grades of GPPS; ②The two prediction models established by PLS-DA and RF can both effectively identify the grades of GPPS. The classification accuracy on the validation set of the PLS-DA model with the optimal principal components of 3 can reach 90.4%. The classification accuracy on the validation set of the RF model with the first five principal components as input variables can reach 95.6%. The RF model shows better grade identification performance than that of the PLS-DA model. Therefore, combined with chemometrics and machine learning algorithms, the in-line near-infrared spectral measurement system can realize the rapid and in-line identification of GPPS grades. It provides a reference for the in-line identification of different grades of the same kind of polymer by near-infrared spectroscopy in a production line.