Machine Learning-Driven Skillful Decadal Predictions of German Bight Storm Surges

Abstract

The German Bight coastline is regularly affected by storm surges driven by extratropical cyclones. Decadal-scale predictions of local surges would foster coastal protection and decision making in affected areas. We examine the prediction skill of the Max-Planck-Institute Earth System Model (MPI-ESM) decadal prediction system for three different storm surge metrics at Cuxhaven (Germany), Esbjerg (Denmark), and Delfzijl (The Netherlands). To avoid dynamical downscaling from the coarse model output to local surge heights, we use machine learning and train a neural network on observed surge heights and reanalyzed fields of mean sea-level pressure (MSLP). We apply this network to MSLP output of our prediction system to generate decadal predictions of surge heights. The prediction system falls short of generating skillful predictions for high water event durations and individual lead years in general, but windows for more skillful predictions arise for deterministic predictions at longer multi-year lead times.