Welcome to the NOAA Institutional Repository |
Stacks Logo
Advanced Search
Select up to three search categories and corresponding keywords using the fields to the right. Refer to the Help section for more detailed instructions.
Clear All Simple Search
Advanced Search
Impr oving NCEP's probabilistic wave height forecasts using neural networks : a pilot study using buoy data
Filetype[PDF-1.38 MB]

  • Description:
    This technical note presents preliminary results (or a pilot study) of neural network models applied to produce non-linear ensemble averaging and bias correction of the Global Wave Ensemble System (GWES) of the US National Weather Service (NWS). Our work seeks to improve the skill of GWES products, including significant wave height (Hs), peak wave period (Tp), and 10-m wind speed from the Global Ensemble Forecast System (GEFS). We present an initial strategy, whereby one location in the Atlantic Ocean and one in the Pacific Ocean, both with reliable and quality controlled buoy data, are used to train and test our statistical models at single points. The GWES was evaluated against National Data Buoy Center (NDBC) measurements at these two points; this comparison indicated an increase of forecast errors and spread with time, as well as an increase of error as a function of percentile levels - indicating the value below which a given percentage of observations in a group of observations fall. Among several tested architectures, the best identified neural network model used two layers, each with 11 neurons at the intermediate layer, a hyperbolic tangent basis function, optimization using sequential training, and normalization applying the log function to time series of Hs. Many different random initializations, with different seeds, were found to have a significant impact on the results. An approach based on an ensemble of neural networks was successfully applied, providing an improvement on the 5-day forecast of 64% in the bias, 29% in the RMSE and scatter index, and 11% in the correlation coefficient. A final neural network model was trained to predict the difference of observations minus the ensemble mean, i.e., the "error" (called residue) of current ensemble average. This approach ensures that no range of values (from calm to extreme events) is deteriorated by the statistical model, and as a consequence expanded the improvement of the neural network post processing to higher percentiles, associated with waves above 2.5 meters. [doi:10.7289/V5/ON-NCEP-490 (http://doi.org/10.7289/V5/ON-NCEP-490)]

  • Document Type:
  • Place as Subject:
  • Main Document Checksum:
  • Supporting Files:
    No Additional Files
No Related Documents.
You May Also Like: