Prediction of influent water quality in wastewater treatment plants based on deep convolutional attention temporal networks

Abstract

[Objective] Under the “dual carbon” goals in China, the accurate prediction of influent water quality in wastewater treatment plants is crucial for energy conservation, emission reduction, and energy consumption reduction. [Methods] To address the insufficient accuracy of traditional influent water quality prediction method(such as artificial neural networks, recurrent neural networks, and long short-term memory networks) in handling the randomness and nonlinearity of wastewater water quality characteristics, a prediction model based on convolutional attention temporal neural network(CAT-NN) was proposed. The model integrated multi-scale information fusion and a hybrid attention mechanism, along with a temporal decoding module, to effectively capture the long-term trends and short-term abrupt changes in wastewater water quality indicators. [Results] Through the predictive analysis of four typical water quality indicators—COD, NH₃-N, TN, and TP—of influent water data from a wastewater treatment plant in Yan'an City, Shaanxi Province, the CAT-NN model demonstrated excellent prediction perfor-mance, with a root mean square error(RMSE) of 4.50% and a mean absolute error(MAE) of 5.00%. Compared to traditional models(such as ANN, LSTM, and gated recurrent units(GRU)), the RMSE and MAE improved by over 16.13% and 20.00%, respectively. [Conclusion] The result indicate that the CAT-NN model achieves higher accuracy and stronger robustness in predicting influent water quality in wastewater treatment plants. The model not only provides strong support for the precise control and efficient operation of wastewater treatment plants, but also serves as a key technological solution for achieving energy conservation and emission reduction goals.

Keywords

wastewater treatment plants / prediction of influent water quality / convolutional attention temporal network / deep learning / carbon neutrality / model performance