Short-term prediction of nitrate plus nitrite in riverine systems using hybrid deep learning models

Nitrate plus nitrite is a key indicator of water quality and ecosystem health. Real-time forecasting remains challenging due to costly in situ sensors and complex nitrogen cycling dynamics. This study developed five deep learning models, CNN, LSTM, TGCN, CNN-LSTM, and TGCN-LSTM, for 7-day-ahead prediction of daily nitrate plus nitrite concentrations. Models were trained on four years of daily data (n = 1461) from two contrasting Chinese rivers: the agriculturally influenced Huangfuchuan River and the urban-impacted Xiangxi River. Inputs combined lagged nitrate measurements with surrogate variables (pH, dissolved oxygen, and water temperature). A sequential 70/30 train-test split preserved temporal causality, supplemented by rolling-origin cross-validation. The TGCN-LSTM hybrid achieved the strongest performance: R² = 0.965 and RMSE = 0.211 mg/L at the agricultural site, and R² = 0.893 and RMSE = 0.499 mg/L at the urban site. Monte Carlo Dropout ensembles provided prediction intervals with 93% and 89% coverage, respectively. Sensitivity analysis confirmed recent nitrate lags as the most influential predictors, while dissolved oxygen and water temperature provided secondary contextual signals; pH contributed minimally. When lagged nitrate was excluded, performance degraded substantially (R² ≈ 0.65-0.75), confirming that past measurements remain essential for high-accuracy forecasting. However, surrogate-only inputs provided operationally useful predictions during sensor outages (R² > 0.70). The model maintained seasonal reliability (R² > 0.87 across all seasons). These findings demonstrate that TGCN-LSTM improves nutrient forecasting accuracy by combining autoregressive nitrate memory with low-cost sensor networks, though operational deployment requires maintaining nitrate sensor continuity or accepting reduced precision during outages.