Decoding Cross-Correlated Hydrochemical Drivers for Robust Nutrient Concentration Prediction

The proper prediction of nutrient concentrations is important for the conservation of river ecosystems and water quality. For this reason, the purpose of this study was to predict daily nitrite and nitrate concentrations at the Sacramento River at Freeport, CA hydrological station on the Sacramento River in the US. Four model types were compared and assessed for their effectiveness, including Clockwork Recurrent Neural Network (CWRNN), Bidirectional Gated Recurrent Unit (BiGRU), a hybrid BiGRU-CWRNN, and an optimal hybrid Red-Billed Blue Magpie Optimizer-BiGRU-CWRNN (RBMO-BiGRU-CWRNN). Model performance was compared under six input scenarios that represented the synchronous, lagged, and cross-correlated effects of hydrological and physicochemical variables. The hybrid approaches performed consistently better than the standalone model approaches, and the RBMO-BiGRU-CWRNN produced the most accurate forecasts. The optimal hybrid model achieved the best predictive skill with R up to 0.899, the lowest RMSE (0.056 mg/l). The NS values of the models were over 0.73 for all models, indicating they had predicted motility with reasonable prediction structures. The models with the short-term lag scenarios produced the best reliability for predictive forecasts of both nitrite and nitrate concentrations, while the longest lags for each nutrient produced the weakest reliability for predictions, demonstrating the challenges longer horizon nutrient residuals present for water quality. Overall, the results confirm that hybrid recurrent networks, along with a metaheuristic optimization approach, can enhance the flexibility and accuracy of model reliability for predicting nutrients. This framework offers a potentially transferable methodological template for tracking nutrient behavior in large river systems, though empirical validation at additional stations and basins is necessary before broader application can be recommended. The approach provides a tool to inform water quality risk assessment within the Sacramento River Basin, with demonstrated robustness at the Freeport Station.