Dual-model LSTM-Transformer framework with hydrodynamic adaptability for TN/TP predictions in tidal reaches

Abstract

The accurate prediction of total nitrogen (TN) and total phosphorus (TP) concentrations is crucial for mitigating eutrophication and supporting sustainable water resource management in tidal estuaries. However, the highly non-stationary dynamics of TN/TP concentrations—characterized by time-varying patterns driven by complex river-tide interactions and biogeochemical cycling—pose significant prediction challenges. While deep learning models like LSTM and Transformer have advanced water quality forecasting, their standalone applications struggle to simultaneously capture long-term dependencies, cross-variable relationships, and hydrodynamic drivers in tidal systems. To address these limitations, we developed a framework that integrates LSTM with Transformer (LSTM-Transformer, LT) and iTransformer (LSTM-iTransformer, LIT), incorporating river discharge (RD), tidal level (TL), rainfall, water temperature (Temp), and PH as key input features. Evaluated across nine sites spanning 600 km of the Yangtze River tidal reach, the dual-model framework demonstrates superior performance over baseline models (LSTM, Transformer, CNN-LSTM), achieving an average RMSE reduction of 32.6 % (range: 16.7–57.1 %). The LT—combining LSTM’s temporal memory with Transformer’s global attention—excels in river-dominated and tide-dominated zones. Meanwhile, the LIT, enhanced through inverted feature embedding, outperforms in transitional zones by resolving short-term variability and cross-parameter interactions. SHAP analysis quantitatively validates hydrodynamic drivers (RD and TL) as critical predictive factors. This study provides a robust modeling solution for TN/TP prediction in tidal reaches, addressing the urgent need for accurate forecasting to enable early pollution warnings and informed water management decisions.