Streamflow Prediction Using Long Shot-Term Memory Models for Optimized Irrigation Management

This study investigates the application of hybrid deep learning architectures—specifically LSTM-GRU models—for daily streamflow forecasting under varying hydrological conditions. Using long-term streamflow and rainfall records from two observation points in a tropical watershed, the models were evaluated across different rolling window lengths and input configurations: streamflow-only, streamflow with rainfall, and streamflow with real-time data assimilation. The results indicate that input type and temporal window selection significantly influence predictive performance. The LSTM-GRU model with real-time data integration consistently achieved the lowest Root Mean Square Error (RMSE ≈ 18.0) and the highest Nash–Sutcliffe Efficiency (NSE > 0.45) for short window sizes (1-year), outperforming both the rainfall-augmented and baseline configurations. Conversely, the model with rainfall input required longer training windows (up to 6 years) to reach optimal accuracy, reflecting delayed hydrological responses. Extending the forecast horizon from one to two years led to a notable decline in model performance across all configurations, though the real-time model remained the most robust and stable. The models also demonstrated strong spatio-temporal transferability, maintaining predictive accuracy across different sub-catchments despite local heterogeneity in watershed behavior. Importantly, integrating LSTM-based forecasts into rotational irrigation systems proved effective in optimizing planting schedules and water allocation, especially during transitional hydrological phases. Overall, the findings underscore the potential of LSTM-GRU models to capture nonlinear, non-stationary hydrological dynamics with high predictive skill. Their capacity to integrate multi-source temporal inputs and adapt to evolving hydrometeorological signals supports more informed, flexible, and resilient water resource planning under climatic uncertainty.