Hybrid modelling for forecasting daily river streamflow in Iran

Accurate streamflow forecasts are essential for flood prediction, reservoir operation, and navigation planning. We develop two hybrid modelling approaches, delta learning (DL) and data augmentation (DA), combining Long Short-Term Memory (LSTM) networks implemented within a nonlinear autoregressive with exogenous inputs (NARX) framework with the Unified Runoff Scheme for Saturation and Infiltration Excess (URSSIE) a process-based hydrological model that simulates saturation and infiltration excess runoff. The framework is applied to forecast 30‑day streamflow for 576 Iranian catchments. DL trains a surrogate to correct residual errors between physical model outputs and observations, whereas DA augments meteorological and streamflow inputs with physical model predictions. Both hybrid methods consistently outperform the standalone hydrological model across all seasons and catchments. Forecast improvements are strongly influenced by hydroclimatic and catchment characteristics, with the greatest gains occurring in northwestern Iran and along the Caspian Sea coasts.