The safety of surface water resources is of great strategic significance. It is related to national health, ecological environment stability and sustainable economic development. Total organic carbon (TOC) is a comprehensive index to reflect the content of organic matter in water. Hence, it has significant value in water environment supervision and treatment. This method is time-consuming and complex. UV-Vis spectroscopy technology has the advantages of fast detection speed and simple operation. Therefore it has a good application prospect in online detection of water quality. At present, the online detection methods of TOC in surface water mostly are indirectly calculated at home and abroad. These methods depend on the correlation between the concentration of COD and TOC, and they require high stability of water composition. Compared with the indirect calculation methods, the spectral quantitative analysis method has better robustness and accuracy. Moreover, this method is convenient for realizing unattended online monitoring of water quality. The experiment was equipped with TOC sample solutions, and a two-day experiment was designed. Six spectral data sets of the samples (denoted as D1, D2, …, D6) were collected in 4 time periods. Firstly, D1 was used as the training set to establish a partial least squares (PLS) regression model in the group experiment. This model was used to predict the TOC concentration of D2, and the mean absolute percentage error (MAPE) was less than 0.78%. In addition, D1 and D2 were collected in the same period. The results show that the established TOC quantitative analysis model has high accuracy. Then, to verify the robustness of the TOC model established by the PLS method to the change of instrument state, the spectral data collected in different periods were selected as the training set, the test set and the validation set. Furthermore, the cross experiments of different instrument states were performed. The MAPE of the predicted TOC concentration in the four experiments were 3.82%, 3.75%, 3.43% and 0.98%, respectively. The results show that the UV-Vis spectroscopy quantitative analysis model of TOC established by the PLS algorithm has good accuracy and robustness. The MAPE of predicted concentration in the grouping experiment and cross experiments of different instrument states are all less than 3.82%. These results are better than the conventional indirect calculation method. Moreover, the established spectral quantitative analysis model does not depend on the calculation relationship between COD and TOC. Thus, it has better adaptability than the conventional indirect calculation method when the water environment changes. Finally, the PLS algorithm has the advantages of a simple modeling process and fast operation speed. It provides convenience for the development and maintenance of submersible online detection equipment.