Improving decadal flood prediction in northern Europe by selecting ensemble members based on North Atlantic Oscillation skill

The ability to predict the frequency and magnitude of flooding at lead times of 1 to 10 years is of great interest to governments and institutions responsible for flood risk management. However, at these lead times there is significant uncertainty about dynamical changes in atmospheric circulation. The current generation of models underestimate the predictable signal of the North Atlantic Oscillation (NAO), the principal mode of variability in North Atlantic atmospheric circulation, leading to low confidence in predictions of regional precipitation and flooding. Recent work has shown that by post-processing a sufficiently large model ensemble, decadal variations in North Atlantic winter climate can become highly predictable (Smith et al., 2020). Here, we investigate whether this NAO-matching technique can be used to improve the skill of flood forecasts at decadal lead times in the United Kingdom. We use a large ensemble of decadal hindcasts consisting of 169 members drawn from CMIP phases 5 and 6, and observed flood records for the period 1960-2015. Following Smith et al. (2020), we adjust the variance of the raw ensemble mean NAO to match that of the observed predictable signal, then select the ensemble members showing the lowest absolute difference with the variance-adjusted ensemble mean. Working only with the selected members (n=20), we supply the ensemble mean precipitation and temperature to a distributional regression model to predict the occurrence and magnitude of winter floods at lead times of 1 to 10 years. We compare these predictions with those from an equivalent model which uses predictors drawn from the full ensemble (n=169) to assess the improvement in predictive skill. Our preliminary results suggest that NAO-matching shows promise at improving decadal flood predictions in northern Europe.

Reference

Smith, D.M., Scaife, A.A., Eade, R. et al. North Atlantic climate far more predictable than models imply. Nature 583, 796–800 (2020). https://doi.org/10.1038/s41586-020-2525-0.