Nonlinear Wave Ensemble Averaging using Neural Networks

Ricardo Campos

UMD/AOSC

20 Feb, Noon, in 2155

20 Feb, Noon, in 2155

This lecture presents results of a GWES assessment using NDBC buoys, studying the errors of 10-m wind speed (U10m), significant wave height (Hs), and peak period (Tp), in function of forecast range and severity (percentiles). Then it focuses on a large experiment using neural networks (NN) applied to nonlinear ensemble averages. First, using a single location approach, considering two buoys in the Pacific and the Atlantic Ocean. Then moving to a spatial approach at the Gulf of Mexico. The NN simulates the residue of the ensemble mean, i.e., the difference from the arithmetic mean of the ensemble members to the buoy observations. The sensitivity NN test considered a total of 12 different numbers of neurons, 8 different filtering windows (residue), and 100 seeds for the random initialization. Independent NN models have been constructed for specific forecast days, from Day 0 to Day 10. Results show that a small number of neurons are sufficient to reduce the bias, while 35 to 50 neurons are optimum to reduce both the scatter and average errors. More complex NN models with a higher number of neurons presented worse results. Finally, a comparison showed significant improvements of the best neural network models (NNs) compared to the traditional arithmetic ensemble mean (EM). The correlation coefficient for forecast Day 10, for example, was increased from 0.39 to 0.61 for U10m, from 0.50 to 0.76 for Hs, and from 0.38 to 0.63 for Tp.