Multisite stochastic simulation under climate change must preserve inter-station dependence while propagating climate signals into basin-scale predictors. Our companion study presented a multisite inflow model for data-sparse regions, evaluated through deterministic recursive simulation, where exogenous climate predictors served to compensate for the short data length available. Here we add a stochastic multisite residual component and present extensions that enable simulation for a future climate. Four exogenous predictors, the East Asian Monsoon Index, Niño3.4, the precipitation-to-potential-evapotranspiration ratio and the Pacific Decadal Oscillation, are conditioned progressively within a multivariate autoregressive base model including interannual dependence, after a spectral transformation that enhances their correspondence with streamflow. Predictors come from six CMIP6 GCMs under SSP2–4.5 over 2025–2064. Monthly affine bias correction is applied to every predictor, Niño3.4 through a two-step ERSSTv5 pipeline that absorbs long-term SST warming rather than propagating a spurious anomaly offset. The rationale is demonstrated at nine dam stations in the Yeong–Seom basin, Republic of Korea, against the 2008–2024 baseline. Driven by bias-corrected GCM predictors over 2008–2024, the full modelling chain reproduces the observed seasonal cycle and, for five of the six GCMs, mean annual inflow to within 7%, and the interannual variance of annual inflow on average, while it underestimates the year-to-year variability of monthly inflow (standard deviation ratio 0.77). Projected mean annual inflows rise across all GCMs and windows by +3% to +47%, with marked seasonal redistribution toward winter and spring. Interannual variance spans 0.65 to 6.8 times observed, with models disagreeing in sign, the dominant uncertainty for multi-decadal planning. Inter-station cross-correlations are maintained to a mean relative RMSE of 8.1%.
Reliable monthly inflow simulation underpins reservoir operation and water resources assessment under a changing climate. This study develops a climate-informed multisite monthly inflow framework based on a residual WAvelet System Prediction (WASP) spectral correction, tested in a system with significant data length constraints. A multivariate autoregressive (MAR) baseline with over-year dependence first represents short-term persistence and concurrent seasonal recurrence across sites. The East Asian Monsoon Index, Niño3.4, the precipitation-to-potential-evapotranspiration ratio (P/PET) and the Pacific Decadal Oscillation are then introduced sequentially, each transformed by WASP against the residual left by the preceding stage rather than the raw inflow, with ridge regression stabilizing estimation. The framework is demonstrated on a monsoon-dominated multi-reservoir basin using monthly records from 2008 to 2024, with 2023–2024 retained for independent deterministic recursive out-of-sample simulation. Averaged across sites, the final stage raised NSE from 0.40 to 0.71, KGE from 0.33 to 0.64, and reduced RMSE from 31.21 to 18.41 MCM. A parallel sequence using untransformed predictors fell below its own baseline at the first two stages and ended 21% higher in RMSE. Because each predictor corrects the preceding residual, the marginal contribution of every climate signal is quantified stage by stage, providing a basis for GCM-informed stochastic inflow ensemble generation, which is developed in a companion study.
Calibrating a distributed rainfall-runoff model is expensive and may depend on expert judgment. Large language model (LLM) agents can now perform the task by reasoning about physical parameters rather than by brute force. Here we ask whether advances in LLM agents carry over to hydrologic model calibration. We ran 26 LLM-harness systems, released between December 2025 and September 2026, on one distributed calibration task. Calibration skill, measured by the Nash–Sutcliffe efficiency (NSE), rises with release date (Spearman rank correlation ρ = +0.56), and catastrophic failures (NSE below zero) become rarer. However, the gain in model calibration does not transfer to independent validation. We re-evaluated every parameter set over the two months after calibration and against the Differential Evolution Adaptive Metropolis (DREAM) sampler given a comparable budget. Compared on trial medians, 12 of 26 systems exceed DREAM during calibration, and none exceeds it during validation. The median system loses 0.46 NSE out of sample, while DREAM loses none. We call on the community to develop a rich and diverse hydrologic evaluation task on out-of-sample datasets to track the progress of frontier LLMs.
Design storms are widely used in hydraulic engineering to represent extreme precipitation and assess flood risk. Commonly used block-method design storms, such as Euler Type II, assign a single return period across all temporal scales of an event, neglecting dependencies in rainfall intensity across durations. This simplification may lead to unrealistic event representations. Here, we compare flood hazard from a traditional Euler Type II storm with storms generated using the multivariate REALISTIC framework, which accounts for joint return periods across durations. Using Zurich as a case study, we simulate 3-hour 100-year events and evaluate the resulting flood hazard with a hydrodynamic model. Results show that flood hazard is highly sensitive to the underlying intensity-duration profile. The block-method storm produces substantially higher flood depths and intensities than events whose profiles are consistent with observed joint return periods, which often result in limited flooding. These findings suggest that traditional design storms provide an incomplete estimation of flood hazard and highlight the need for multivariate approaches to support more robust risk assessments.
