Short-Term Prediction and Temporal Causality Analysis of Total Nitrogen in Wastewater Treatment Plant Effluent Based on LT-PR-LSTM

Accurate prediction of effluent total nitrogen (TN) is important for early exceedance warning and operational control in wastewater treatment plants. Existing decomposition-based models may overestimate performance when full-series decomposition is performed before data splitting, causing potential temporal information leakage. To address this issue, this study compares noncausal and strictly causal Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise combined with Long Short-Term Memory (ICEEMDAN-LSTM) and Variational Mode Decomposition–Long Short-Term Memory network (VMD-LSTM) settings, and proposes a Level–Trend Persistence-Residual LSTM (LT-PR-LSTM) for univariate effluent TN prediction. The model uses Persistence as the short-term state baseline, extracts level features from historical TN, and introduces first- and second-order differences to learn residual corrections relative to the current state. Multi-model comparison, ablation experiments, stability tests, SHapley Additive exPlanations (SHAP) interpretation, supplementary dataset validation, and efficiency analysis were conducted. Results show that noncausal decomposition inflates predictive performance. LT-PR-LSTM achieves the best main-test performance, with RMSE 1.1273, MAE 0.6082, MAPE 7.5455%, and R2 0.8512, reducing RMSE, MAE, and MAPE by 6.73%, 7.64%, and 8.56% compared with Persistence. SHAP identifies TN(t−2h) as the dominant predictor, and the model requires only 0.5348 ms/sample, indicating potential for online TN early warning.