ABSTRACT Understanding watershed water quality dynamics is essential for sustainable management, yet accurate nutrient load prediction remains challenging under strong inter‐annual variability. To address this limitation, this study presents a hybrid modelling framework that integrates baseflow information into a machine‐learning structure to improve nutrient load prediction. By separating and quantifying baseflow contributions, the proposed approach provides a process‐informed foundation for data‐driven prediction. We employed a conventional Long Short‐Term Memory (LSTM) model as a baseline and developed a hybrid model incorporating baseflow nutrient load contribution. In addition, the limitations of applying conventional environmental models in watersheds with strong seasonality were explored. The results show that the hybrid approach significantly outperformed the standard LSTM and process‐based models. The benchmark LSTM model exhibited a percentage bias (PBIAS) of −3.08% to −126.57% and a Nash‐Sutcliffe Efficiency (NSE) of 0.13–0.95. The hybrid models reduced PBIAS to −1.88% to 47.21% and increased NSE to 0.66–0.99. Notably, this improvement was pronounced during wet seasons, indicating that incorporating baseflow information strengthens prediction accuracy at peak flow conditions. These findings demonstrate that accounting for baseflow contributions enhances nutrient load prediction in machine‐learning frameworks, particularly in watersheds with high hydrological variability.
Yifru et al. (Wed,) studied this question.