Global and coastal ocean surface water elevation prediction skill has advanced considerably with improved algorithms, more refined discretizations and high-performance parallel computing. Model skill is tied to mesh resolution, the accuracy of specified bathymetry/topography, dissipation parameterizations, air-sea drag formulations, and the fidelity of forcing functions. Wind forcing skill can be particularly prone to errors, especially at the land-ocean interface. The resulting biases and errors can be addressed holistically with a machine-learning (ML) approach. Herein, we weakly couple the Temporal Fusion Transformer to the National Oceanic and Atmospheric Administration’s (NOAA) Storm and Tide Operational Forecast System (STOFS 2D Global) to improve its forecasting skill throughout a 7-day horizon. We demonstrate the transformer’s ability to enrich the hydrodynamic model’s output at 228 observed water level stations operated by NOAA’s National Ocean Service. We conclude that the transformer is a rapid way to correct STOFS 2D Global forecasted water levels provided that sufficient covariates are supplied. For stations in wind-dominant areas, we demonstrate that including past and future wind-speed covariates make for a more skillful forecast. In general, while the transformer renders consistent corrections at both tidally and wind-dominant stations, it does so most aggressively at tidally-dominant stations. We show notable improvements in Alaska and the Atlantic and Pacific seaboards of the United States. We evaluate several transformers instantiated with different hyperparameters, covariates, and training data to provide guidance on how to enhance performance.
