System-level total nitrogen prediction in MLE wastewater treatment using a modular merged LSTM framework

Total nitrogen (TN) prediction in full-scale wastewater treatment plants remains challenging due to nonlinear inter-stage interactions and variable operating conditions. Conventional LSTM or hybrid deep-learning models often treat the process as a single sequence, limiting system-level interpretability and generalization. This study develops a Modular Merged LSTM (MM-LSTM) framework that explicitly decomposes the Modified Ludzack–Ettinger (MLE) process into three interconnected modules—influent (IN), primary sedimentation (1st SD), and bioreactor with secondary sedimentation (BIO–2nd SD)—which are sequentially merged through a trainable output-level fusion layer. This modular design captures cross-stage dependencies while maintaining process-specific heterogeneity. Using five-year daily monitoring data (2019–2023) from a full-scale MLE line, MM-LSTM was benchmarked against (1) Local-LSTM (stage-only inputs) and (2) Upstream-Aware LSTM (target + upstream inputs). The MM-LSTM achieved NSE = 0.667, RMSE = 2.48 mg/L, and MAE = 1.79 mg/L, representing 7 % improvement in NSE and 3 % reduction in RMSE and MAE compared with UA-LSTM, with the smallest train–test degradation. Interpretability was enhanced by analyzing (a) module-level contributions via merging-layer weights and (b) feature-level attributions via Integrated Gradients, which consistently identified TN, aeration, and internal recycle as dominant predictors. The proposed framework is scalable, interpretable, and transferable, offering a robust basis for system-level TN prediction, early-warning applications, and adaptive aeration control under dynamic operating conditions.