Non-point source (NPS) pollution is an important factor affecting the quality of water environment. In recent years, a large number of online water quality monitoring stations have been used to obtain continuous time series water quality monitoring data. These data provide the necessary basis for the application of deep learning methods in water quality prediction. However, the prediction accuracy of traditional deep learning methods is low, especially in predicting the water quality with NPS pollution. Aiming to address this limitation, a novel deep learning model named SOD-VGG-LSTM with the simulation-observation difference (SOD) modular based on physical process, the visual geometry (VGG) modular reflecting spatial characteristics, and the long short-term memory (LSTM) modular based on deep learning method was developed to improve the accuracy of the water quality prediction with NPS pollution. The established model can overcome the problem that mechanism models can not predict the changes of water quality on the hourly or minute time scale. The model was applied in Lijiang River watershed. Experimental results indicated that the proposed model had the highest accuracy in the extreme value prediction compared with the mechanism model and LSTM model. The maximum relative errors between the predicted and observed results for DO, COD_Mn, NH₃-N, and TP were 8.47%, 19.76%, 24.1%, and 35.4%, respectively. The model evaluation demonstrated that the established SOD-VGG-LSTM model achieved superior computational performance compared to Auto Regression Integreate Moving Average model (ARIMA), Support Vector Regression model (SVR), and Recurrent Neural Network model (RNN). The evaluation results showed that SOD-VGG-LSTM achieved 3.2–39.3% higher R² than ARIMA, SVR and RNN. The proposed model can provide a new method for water quality prediction with NPS pollution.

非点源（NPS）污染是影响水环境质量的重要因素。近年来，大量的在线水质监测站被用于获取连续的时间序列水质监测数据。这些数据为深度学习方法在水质预测中的应用提供了必要的基础。然而，传统深度学习方法的预测精度较低，尤其是在预测具有 NPS 污染的水质方面。为了解决这一限制，一种名为 SOD-VGG-LSTM 的新型深度学习模型具有基于物理过程的模拟观察差异 (SOD) 模块、反映空间特征的视觉几何 (VGG) 模块，并开发了基于深度学习方法的长短期记忆（LSTM）模块，以提高NPS污染水质预测的准确性。所建立的模型可以克服机理模型无法预测每小时或分钟时间尺度上水质变化的问题。该模型应用于漓江流域。实验结果表明，与机理模型和LSTM模型相比，该模型在极值预测方面的准确率最高。DO、COD 的预测结果和观测结果之间的最大相对误差 该模型应用于漓江流域。实验结果表明，与机理模型和LSTM模型相比，该模型在极值预测方面的准确率最高。DO、COD 的预测结果和观测结果之间的最大相对误差 该模型应用于漓江流域。实验结果表明，与机理模型和LSTM模型相比，该模型在极值预测方面的准确率最高。DO、COD 的预测结果和观测结果之间的最大相对误差_Mn、NH ₃ -N 和TP 分别为8.47%、19.76%、24.1%和35.4%。模型评估表明， 与自动回归集成移动平均模型（ARIMA）、支持向量回归模型（SVR）和循环神经网络模型（RNN）相比，所建立的SOD-VGG-LSTM模型具有更好的计算性能。^{评估结果表明，SOD-VGG-LSTM 的 R 2}比 ARIMA、SVR 和 RNN高 3.2-39.3% 。该模型可为NPS污染水质预测提供一种新方法。