A Graph Attention-Enhanced Hybrid Deep Learning Model for Effluent Total Nitrogen and Total Phosphorus Prediction in Municipal WWTPs

Accurate effluent-quality prediction is essential for improving nitrogen and phosphorus removal performance and reducing energy consumption in wastewater treatment plants (WWTPs). However, the strong coupling, high noise, and time-lag effects in wastewater treatment processes pose significant challenges to existing prediction models. In this study, we propose a GAT-CNN-LSTM(GCL) model for the prediction of effluent total nitrogen (TN) and total phosphorus (TP). The GCL model first uses a graph attention network (GAT) to adaptively learn inter-variable relationships, and then applies a convolutional neural network (CNN) and long short-term memory (LSTM) network to extract local and long-term temporal features. The GCL model is trained and evaluated using real operational data from a municipal WWTP in northern China. Based on the best run of each model, GCL improves the R2 by 13.7% and 6.4% over LSTM and Transformer for TN prediction, while reducing MAPE by 39.4% and 30.4%, respectively. For TP prediction, the corresponding improvements in R2 are 70.7% and 59.1%, with MAPE reductions of 37.1% and 36.0%. Ablation experiments further demonstrate the complementary contributions of the three modules, showing that graph-based feature fusion enhances subsequent temporal modeling. The temporal variation in neighbor attention weights and one-at-a-time (OAT) sensitivity analysis provide interpretability consistent with A2/O process mechanisms. These findings provide a preliminary validation based on a limited dataset from a single WWTP, and broader applicability under more diverse operating conditions warrants further investigation.