Prediction model of drinking water source quality with potential industrial-agricultural pollution based on CNN-GRU-Attention

It is of great significance for the operation and management of waterworks to accurately predict the raw water quality. In this study, the turbidity and CODcr of raw water with light industrial-agricultural pollution are predicted based on the long-term water quality data of Bayi Waterworks in Xiaogan. In order to solve the problem that the traditional hydraulic models require excessive calculation with relatively poor prediction results, this paper presents a CNN-GRU-Attention (CGA) model to predict water quality. In CGA model, CNN layer can extract the short-term features of water quality data, GRU model can extract the long-term features, and attention layer can adjust the weights of different neurons according to the correlation among neurons, to increase their interconnections. We compare the CGA model with individual LSTM and GRU models, and find that the combined model is better than the single model, and the lower the data correlation is, the better the optimization effect is. When CGA model is compared with CNN-LSTM-Attention model, it is found that CGA model has a little improvement in prediction effect and a great improvement in model stability. In order to evaluate the quality of water source comprehensively, an improved TOPSIS model is established by using correlation relation for weight. The results show that the water quality is slightly worse in summer than in winter, and the overall raw water quality currently has seen a gradual downturn due to the local industrial-agricultural pollution in the neighboring area. Finally, we collectively evaluate the composition of main pollutants in the raw water and the performance of water treatment units to deliberate on the results of water quality prediction, in an effort to provide a decision-making basis for water quality integrated management and pollutant control, which should put a priority on controlling organic substances and NH3-N, in a typical minor-polluted river basin in Central China.

Automated summary

Accurately predicting the quality of raw water is of great significance for the operation and management of waterworks. In this study, a CGA model was proposed to predict the turbidity and CODcr of raw water with light industrial-agricultural pollution, based on long-term water quality data. The CGA model, which combines CNN, GRU, and attention layers, outperforms individual LSTM and GRU models. The study findings also indicate a gradual decline in the overall raw water quality due to local industrial-agricultural pollution.