Integrating Hydrological-Hydraulic-AI (LSTM) Models for Improved Water Level Forecasting: Red River - Thai Binh Basin

Abstract

Flooding remains one of the most severe natural hazards in Vietnam, causing recurrent damage to infrastructure, livelihoods, and socio-economic development. The Red River - Thai Binh Basin, a densely populated and economically vital region, is particularly vulnerable due to its complex hydrological regime, tidal influences, and rapid urbanization. Improving water level forecasting in this basin is therefore critical for effective flood risk management and disaster preparedness. This study develops an integrated decision-support tool that combines three complementary modelling approaches: the NAM rainfall-runoff model, the MIKE 11 hydraulic model, and a Long Short-Term Memory (LSTM) neural network. Hydrometeorological data from 34 meteorological and 5 hydrological stations, together with topographic (DEM 30 × 30 m) and land cover datasets, were used for model setup, calibration, and validation. The models were calibrated with 2023 data and validated with 2024 data, while two extreme flood events, Typhoon Wipha July 2025 and Typhoon Kajiki in August 2025 were applied for real-time testing. Results show that the integrated framework achieves high predictive accuracy, with Nash-Sutcliffe Efficiency (NSE) values up to 0.97 and Root Mean Squared Error (RMSE) generally below 0.25 m across stations. The Long Short-Term Memory (LSTM) component significantly improved forecasts in downstream areas where tidal effects challenge conventional models. The proposed tool demonstrates strong potential for operational use, offering timely and reliable water level forecasts to support flood warning and management. Beyond the Red River - Thai Binh Basin, the framework is replicable for other flood-prone river systems in Southeast Asia, providing both methodological and practical contributions to climate resilience and disaster risk reduction.