Probabilistic predictions of storm surge from coarse scale climate data based on normalizing flows

Storm surges are causing widespread devastation, directly impacting coastal communities through injuries and fatalities, infrastructure damage, and the displacement of residents. Projections of future storm surges are vital for assessing these risks, especially under climate change that causes both the intensity and frequency of these extreme events to increase. The temporal and spatial resolution of global climate model simulations do not resolve the critical characteristics of the events: storm surge peaks such as daily maximum storm surge occur on the scale of hours, while global climate model simulations are often only available at daily time scales. The coarse resolution data include some information about the daily maximum water levels but does not exactly determine the maximum storm surge peaks. Instead, a range of daily maximum storm surge peaks are realistic under the same coarse conditions.

Hydrodynamic and data-driven models often derive storm surge time series deterministically capturing the average outcome, but do not represent the range of outcomes given coarse-scale predictors. Probabilistic models can address this by generating ensembles of outcomes, each consistent with coarse-scale predictors. For future projections, where no observed storm surge exists for comparison, it can be beneficial to use individual ensemble members to provide more realistic storm surge scenarios.

We implement a multivariate probabilistic model using normalizing flows to simulate time series ensembles of daily maximum storm surges, driven by climate data aggregated to daily means. We train and evaluate the model using ERA5 climate reanalysis data and storm surge time series from the hydrodynamic Global Tide and Surge Model in the time period 1979-2018 across five representative regions worldwide. Our findings indicate that individual ensemble members replicate key statistical features of storm surges more effectively than the ensemble means, given the limited temporal and spatial resolution of the predictors. The multivariate model effectively preserves spatial correlations within each individual ensemble member, making it a spatially realistic realization of storm surge. 

Probabilistic storm surge time series conditioned on coarse atmospheric predictors open up new possibilities beyond traditional hydrodynamic modeling. Its performance in settings with limited predictor resolution make it an effective tool for computing storm surge projections consistent with climate model outputs.