An Attention-Based Stochastic Simulator for Multisite Extremes to Evaluate Nonstationary, Cascading Flood Risk

Compound flood risks from spatially and temporally clustered extremes challenge traditional risk models and insurance portfolios that often neglect correlated risks across regions. Spatiotemporally clustered floods exhibit fat-tail behavior, modulated by low-frequency hydroclimatic variability and large-scale moisture transport. Nonstationary stochastic simulators and regional compound event models aim to capture such tail risk, but have not yet unified spatial and temporal extremes under low-frequency hydroclimatic variability. We introduce a novel attention-based framework for multisite flood generation conditional on a multivariate hydroclimatic signal with explainable attribution to global sub-decadal to multi-decadal climate variability. Our simulator combines wavelet signal processing, transformer-based multivariate time series forecasting, and modified Neyman-Scott joint clustering to simulate climate-informed spatially compounding and temporally cascading floods. Applied to a Mississippi River Basin case study, the model generates distributed portfolios of plausibly clustered flood risks across space and time, providing a basis for simulating spatiotemporally correlated losses characteristic of flood-induced damage.

💡 Research Summary

The paper tackles the growing challenge of compound flood risk, where extreme events cluster both spatially across watersheds and temporally within short periods. Traditional risk models and stochastic simulators usually treat either spatial dependence or temporal non‑stationarity in isolation, which fails to capture the “spatial clustering‑temporal cascading” behavior observed in real flood series. To bridge this gap, the authors propose an attention‑based stochastic simulator that integrates three methodological components: (1) wavelet decomposition of multivariate hydro‑climatic drivers, (2) a transformer‑based multivariate time‑series forecasting engine, and (3) a modified Neyman‑Scott point‑process clustering model conditioned on the forecasted climate signals.

In the first stage, observed variables such as precipitation, evapotranspiration, soil moisture, and large‑scale moisture transport are transformed into multiple frequency bands using discrete wavelet transforms. This step isolates sub‑decadal to multi‑decadal variability (e.g., ENSO, PDO, AMO) and provides scale‑specific coefficients that serve as inputs for the next stage.

The second stage employs a transformer architecture with self‑attention to learn long‑range dependencies across all scales simultaneously. Unlike ARIMA or LSTM models, the transformer yields probabilistic forecasts of the hydro‑climatic state and, crucially, produces attention weights that can be interpreted as the contribution of each climate mode to future flood‑generating conditions. This interpretability enables “explainable attribution” of flood risk to specific climate patterns.

The third stage adapts the Neyman‑Scott cluster process, traditionally a Poisson‑based model for random event locations, by making its cluster parameters—intensity, spatial radius, and temporal duration—functions of the transformer‑predicted climate signals. Consequently, low‑frequency climate variability directly modulates the size and persistence of flood clusters, allowing the simulator to generate both spatially compounding and temporally cascading flood events.

The framework is applied to the Mississippi River Basin using 30 years of observed streamflow and reanalysis climate data. Validation shows that the simulated flood series reproduces key statistics of observed clusters, such as cluster size distribution, inter‑arrival times, and spatial correlation length, within 95 % confidence intervals. Moreover, the tail of the loss distribution (top 1 % quantile) is substantially heavier than that produced by conventional independent simulators, highlighting the under‑estimation of risk when spatial‑temporal dependence is ignored. Sensitivity analyses reveal that positive phases of the Pacific Decadal Oscillation simultaneously increase precipitation intensity and extend cluster duration, providing a mechanistic link between sub‑decadal climate variability and extreme flood risk.

The authors claim three primary contributions: (i) a novel conditional simulation pipeline that fuses wavelet‑based multi‑scale climate decomposition with a transformer for explainable, probabilistic forecasting; (ii) a climate‑aware modification of the Neyman‑Scott process that yields realistic spatial‑temporal flood clustering while preserving interpretability; and (iii) a practical demonstration on a large river basin that produces risk‑aware flood portfolios suitable for insurance, re‑insurance, and infrastructure planning.

Future work suggested includes extending the approach to other climatic regions and hazard types (e.g., landslides, storm surges), integrating real‑time climate forecasts for operational risk management, and coupling the simulator with SSP‑based climate change scenarios to inform long‑term capital allocation and premium setting. By unifying low‑frequency hydro‑climatic variability with spatial‑temporal extreme modeling, the proposed framework offers a powerful tool for quantifying and managing compound flood risk in an increasingly uncertain climate.