Hybrid process-based and deep learning for river nutrient prediction under limited monitoring data

• Developed deep learning long short-term memory (LSTM) methods for nutrient simulation in a data-limited catchment. • Benchmarked observation-based (empirical) LSTM models against hybrid models informed by process-based simulations. • Identified impacts of extreme weather indices: improved predictions of phosphorus but reduced performance for nitrogen. • Revealed interactions between LSTM memory and process-based model outputs, showing that pre-river nutrient loads provide stronger predictive signals than in-river concentration residuals. Accurate simulation of riverine nutrient dynamics remains challenging in catchments with limited monitoring data, where both process-based models (PBMs) and data-driven approaches face constraints. This study investigates hybrid modelling strategies that combine PBM simulations with long short-term memory (LSTM) networks to predict nitrogen and phosphorus concentrations at the outlet of Salmons Brook catchment, London, UK. We designed two scenario sets to evaluate (i) the role of extreme weather indices as input features and (ii) the value of integrating different outputs from the Water Systems Integrated Modelling framework (WSIMOD) into LSTM architectures. Results show that the selected hybrid PBM-LSTM models outperformed both empirical LSTM and standalone WSIMOD simulations. For phosphorus, incorporating extreme weather indices improved performance, reflecting its sensitivity to high air temperatures and intense precipitation, while nitrogen predictions degraded, suggesting that memory from past water quality observations carries more predictive value. Comparisons between hybrid designs further indicate that simulated pre-river nutrient loads provide stronger constraints for LSTM models than in-river nutrient processes. These findings emphasize the need for nutrient-specific observational inputs in LSTM frameworks and demonstrate that PBM-LSTM hybridization offers a promising pathway for improving predictions relative to standalone LSTM or PBM simulations. Moreover, the comparative performance of different modelling configurations offers insights into the types of signals LSTM networks retain or discard, thereby contributing to the interpretability of DL applications in river water quality prediction.