A lag-informed deep learning framework for real-time downstream water-level forecasting and flood decision support in the Kalu Ganga Basin, Sri Lanka

Floods are among the most damaging hazards in Sri Lanka's wet-zone river basins, where short warning lead times constrain preparedness and physically based hydrodynamic models are often impractical under limited data. This study develops a lag-informed deep-learning framework for 10-h-ahead downstream water-level forecasting in the flood-prone Kalu Ganga Basin, Sri Lanka. Flood-wave travel times between Rathnapura, Ellagawa, Millakanda, and the downstream target station Putupaula were estimated from hourly records covering 2021–2024 using event-based peak matching, rising-limb timing, and cross-correlation. Selected travel times were 18 h for Rathnapura–Ellagawa, 23 h for Ellagawa–Putupaula, and 24 h for Millakanda–Putupaula. These lags were used to align upstream water-level inputs, while rainfall remained in its original chronological order. No-lag and lag-informed Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Bi-LSTM models were evaluated using chronological training, validation, and independent testing. Lag-informed inputs reduced Root Mean Square Error (RMSE) from 0.317 to 0.288 m for LSTM and from 0.339 to 0.286 m for GRU, while Bi-LSTM performance decreased. The lag-informed GRU also outperformed persistence and ARIMA baselines. Stage-aware training further improved high-flow prediction, with the GRU achieving an RMSE of 0.262 m, Mean Absolute Error (MAE) of 0.188 m, and Nash–Sutcliffe Efficiency (NSE) of 0.934, while the LSTM predicted the 5.57 m June 2024 flood peak at 5.44 m. The framework was deployed as a real-time decision-support dashboard named KaluHydroNet, demonstrating that lag-informed and stage-aware deep learning coupled with live data integration offers a practical pathway for short-term flood early warning in the Kalu Ganga Basin.