Application of LSTM coupled models in runoff simulation and prediction: a review

Accurate runoff simulation and prediction are critical for effective water resource management. While traditional physical models incorporate hydrological mechanisms, they often struggle with nonlinearity and uncertainty. Conversely, pure data-driven models such as Long Short-Term Memory (LSTM) networks excel at capturing temporal patterns but lack physical constraints. This review systematically examines the emerging paradigm of coupled LSTM models, which integrate the strengths of both approaches to boost predictive accuracy and stability. This review categorizes and analyzes three dominant coupling frameworks: the integration of LSTM with physical models to leverage physical insights for improved generalization; the combination of LSTM with signal decomposition techniques to disentangle complex hydrological signals for more accurate forecasting; and the coupling of LSTM with intelligent optimization algorithms to automate feature selection and hyperparameter tuning. Analysis reveals that these coupled models significantly outperform standalone methods across various case studies. The core findings indicate that coupling enhances model accuracy, provides insights into hydrological processes, and offers solutions for data-scarce regions through physical priors. Finally, current challenges and future directions are discussed, highlighting the potential of these integrated approaches to advance hydrological sciences and support sustainable water management decisions.