Abstract: With the development of science and technology, numerous online monitoring points for water quality and meteorology have been established in key river basins. The acquisition of long-term water quality and meteorological data can be acquired and used to train deep learning models, which is the basis for the simulation and prediction of river water pollution. A multi-mode hybrid model was proposed to feature extraction and time-series prediction. In order to simulate the changing process of water pollution under real meteorological conditions, the changing features of water quality were extracted by VGG (Visual Geometric Group), and time-series data was trained by GRU (Gated Recurrent Unit). For the river pollution of Lijiang River, an internationally renowned eco-tourism area, using the online monitoring data of water quality and related meteorology at 47 stations, the proposed model was applied and evaluated. At different time intervals, the experimental results were compared with multiple advanced methods. The experimental results showed that the proposed WQ-VGG-GRU model could effectively improve the accuracy of water quality prediction. Compared with other methods, the mean absolute error (MAE), root mean square error (RMSE), and average absolute percentage error (MAPE) of the predicted values were increased by 8.6%, 13.8%, and 13.7%, respectively, which showed that the proposed model was more accurate in predicting extreme values and had better applicability.