This study aimed at (i) developing a multi-station deep learning rainfall–runoff–operation (DL-RRO) model for cascading dam systems and (ii) applying the model through the large ensemble dataset for Policy Decision-Making for Future Climate Change (d4PDF) to assess flood characteristic changes. The DL-RRO model was developed using a Long Short-Term Memory network that coupled rainfall–runoff processes and mimicked existing dam operation rules by learning operational logic from observed data. Dynamic water level–storage curves were generated to account for time-varying increases in sediment storage. The proposed method was applied to the upper Chikugo River Basin (CRB), where the Shimouke and Matsubara Dams play a critical role in flood control in Kyushu, Japan. The model showed high performance when trained, validated, and tested with 40 years (1986–2025) of hourly hydrological observations (NSE = 0.89, RMSE = 18.90 m 3 /s). Subsequently, d4PDF rainfall data for the future scenarios (2031–2090 and 2051–2110; 1,440 ensemble-year) were imposed to the DL-RRO model. Analysis of 8,725 simulated flood events revealed increased flood frequency and notable changes in hydrograph characteristics. Projected 100-year return flood may increase by +53 % at Shimouke and +165 % at Matsubara under future scenarios, exceeding existing spillway design capacities. Under the synergistic impacts of climate change and sediment dynamics, amplified flood characteristics may intensify flood-driven sediment deposition, leading to further reductions in flood control capacity of up to +2.9 % at the Shimouke and +4.1 % at the Matsubara. These findings highlighted the urgent need to upgrade dams, revise operations and sedimentation management strategies in the upper CRB.
NGUYEN et al. (Wed,) studied this question.