In response to increasing flood risks driven by the climate crisis, urban areas require advanced forecasting and informed decision-making to support sustainable development. This study seeks to improve the reliability of reservoir-based flood forecasting and ensure adequate lead time for effective response measures. The main objectives are to predict hourly downstream flood discharge at a reference point, compare discharge predictions from a single reservoir with a four-hour lead time against those from three reservoirs with a seven-hour lead time, and evaluate the accuracy of data-driven approaches. The study takes place in the Han River Basin, located in Seoul, South Korea. Approaches include two non-deep learning (NDL) (random forest (RF), support vector regression (SVR)) and two deep learning (DL) (long short-term memory (LSTM), gated recurrent unit (GRU)). Scenario 1 incorporates data from three reservoirs, while Scenario 2 focuses solely on Paldang reservoir. Results show that RF performed 4.03% (in R2) better than SVR, while GRU performed 4.69% (in R2) better than LSTM in Scenario 1. In Scenario 2, none of the models showed any outstanding performance. Based on these findings, we propose a two-step reservoir-based approach: Initial predictions should utilize models for three upstream reservoirs with long lead time, while closer to the event, the model should focus on a single reservoir with more accurate prediction. This work stands as a significant contribution, making accurate and well-timed predictions for the local administrations to issue flood warnings and execute evacuations to mitigate flood damage and casualties in urban areas.
