Improving short-term wave forecasts with convolution neural networks and regional buoy observations

Jonny Z Mooneyham, Western Washington University, Computer Science, Bellingham, United States, Sean C Crosby, Western Washington University, Geology, Bellingham, WA, United States, Nirnimesh Kumar, University of Washington, Seattle, WA, United States and Brian Hutchinson, Western Washington University, Computer Science, Bellingham, WA, United States; Pacific Northwest National Laboratory, Computing & Analytics Division, Richland, United States
Accurate nearshore wave forecasts are needed to improve public and boating safety; and for timely alerts of high waves, strong alongshore currents, extreme run-up, and increased rip-current likelihood. Though bulk wave statistics, such as wave height and period, are the focus of most forecasts, wave spectra and directional details are often critical in sheltered regions and for local hydrodynamic models where incident wave energy flux gradients set up alongshore and cross-shore currents. Though global operational wave model skill continues to improve, assimilation of spectral details is difficult and current operational models do not assimilate buoy observations. Until seamless algorithms for assimilation of regional buoy observations are developed, there are opportunities to post-process model predictions with corrections developed by recent observation-prediction mismatch. Prior work has shown skill in generating predictions from recent observations using machine learning and neural networks, but have focused primarily on bulk wave properties. Here, a convolutional neural network (CNN) is trained to make forecast corrections from recent Wave Watch III (WW3) predictions and directional wave buoy observations. Frequency-directional WW3 spectra at offshore buoy locations, transformed into directional buoy moments, are simultaneously corrected across all frequencies. The CNN is trained with 10-years of overlapping NOAA's WW3 CFSR Reanalysis phase 2 predictions and buoy observations at three locations offshore of California and Washington state, USA. Lacking a spectral forecast archive, reanalysis predictions are used to create synthetic forecasts. The trained CNN model reduces wave height RMSE by 10-50% during the first 6-hours, and 10-20% thereafter. Skill improvement is evident up to 24-hours, though error reduction between 12-24 hours is marginal. Errors are reduced across frequency, with most improvement for seas (5-10 seconds), and least for low-frequency swell (12-20 seconds). Mean direction RMSE is reduced by 20-50%, and additionally bias is lowered for alongshore radiation stress predictions (typically used to infer alongshore transports). CNN models trained on Washington buoys are skillful in California, though less so compared to site-specific models. The skill observed using models interchangeably across location implies results are generalizable, and that future training sets may be composites from varying locations.