Enhanced Time Series–Physics Model Approach for Dam Discharge Impacts on River Levels: Seomjin River, South Korea

In dam operations, sudden discharges during extreme rainfall events can pose severe flood risks to downstream communities. This study developed a dam discharge-based river water level forecasting model using a data-driven deep learning approach, long short-term memory (LSTM). To enhance predictive performance, physics-based HEC-RAS simulation outputs, including extreme events, were incorporated as additional inputs. The Seomjin River Basin in South Korea, which recently experienced severe flooding, was selected as the study area. Hydrological data from 2010 to 2023 were utilized, with 2023 reserved for model testing. Forecasts were generated for four lead times (3, 6, 12, and 24 h), consistent with the operational flood forecasting system of the Ministry of Environment, South Korea. Using only observed data, the model achieved high accuracy at upstream sites, such as Imsil-gun (Iljung-ri, R2 = 0.92, RMSE = 0.27 m) and Gokseong (Geumgok Bridge, R2 = 0.91, RMSE = 0.35 m), for a 6-h lead time. However, performance was lower at Gurye-gun (Songjeong-ri, R2 = 0.72, RMSE = 1.48 m) due to the complex influence of two dams. Incorporating enhanced inputs significantly improved predictions at Gurye-gun (R2 = 0.91, RMSE = 1.17 m at 3 h). Overall, models using only observed data performed better at upstream sites, while enhanced inputs were more effective in downstream or multi-dam regions. The 6-h lead time yielded the highest overall accuracy, highlighting the potential of this approach to improve real-time dam operations and flood risk management.