A Stochastic Machine-Learning Based Approach for Hydrologic Drought Risk Characterization in Taiwan

Taiwan’s water security is highly dependent on seasonal rainfall, particularly typhoons and monsoon rains that replenish its more than 100 reservoirs each year. When less rain falls than is expected, droughts can occur that threaten agriculture, industry, and domestic water supply. To assess future drought risk over the next two decades, we develop a stochastic reservoir modelling framework that links meteorological variability to reservoir storage through machine learning. Using historical data from five reservoirs in southern Taiwan, we train an extreme gradient boosted trees (XGBoost) model to simulate weekly reservoir storage changes as a function of precipitation and storage history. Stochastic precipitation sequences are generated from both historical observations and CMIP6 climate projections to produce ensembles of synthetic reservoir storage time series. Validation shows that the model captures key seasonal dynamics and drought characteristics for some of the reservoirs, though it tends to underestimate drought severity and duration compared to the observational record. Projections for the 2040s suggest that droughts may occur more frequently and with greater variability in onset timing, but with shorter durations and reduced severity relative to historical extremes. These results highlight emerging challenges for reservoir management under climate change, particularly the need to adapt operations to increasing uncertainty in drought timing and frequency. Our framework demonstrates the utility and limitations of combining stochastic hydrology and machine learning for reservoir modelling and future drought assessment.