The US Navy's RELO ensemble prediction system and its performance in the Gulf of Mexico

Abstract

Listen

Translate article

AbstractThe US Navy's relocatable (RELO) ensemble prediction system is fully described and is examined in the Gulf of Mexico for 2010. After briefly describing the ensemble transfer (ET) method for the initial perturbation generation, we introduce a new time‐deformation technique to generate the surface forcing perturbations from the atmospheric model fields. The extended forecast time (EFT) is introduced to quantify the advantages of the ensemble mean forecasts over a single deterministic forecast. The ensemble spread and its growth are investigated together with their relations with the ensemble forecast accuracy, reliability and skill.Similar to many other operational ensemble forecast systems at numerical weather prediction (NWP) centres, the initial analysis error is underestimated by the technique used in the data assimilation (DA) system. Growth of the ocean ensemble spread is also found to lag the growth of the ensemble mean error, a tendency attributed to insufficiently accounting for model‐related uncertainties. As an initial step, we randomly perturb the two most important parameters in the ocean model mixing parametrizations, namely the Smagorinsky horizontal and Mellor–Yamada vertical mixing schemes. We examine three different parameter perturbation schemes based on both uniform and Gaussian distributions. It is found that all three schemes improve the ensemble spread to a certain extent, particularly the scheme with Gaussian distribution of perturbations imposed on both the horizontal and vertical mixing parameters.The findings in this article indicate that the RELO ensemble forecast demonstrates superior accuracy and skill relative to a single deterministic forecast for all the variables and over all the domains considered here. The ensemble spread provides a valuable estimate of forecast uncertainty. However, the RELO uncertainty forecast capability could be further improved by accounting for more model‐related uncertainties, for example, by the development of an error parametrization that imposes stochastic forcing at each model grid point.