Operational surface drift prediction using linear and non-linear hyper-ensemble statistics on atmospheric and ocean models

Abstract

Multimodel super-ensemble forecasts, which exploit the power of an optimal local combination of individual models usually show superior forecasting skills when compared to individual models because they allow for local correction and/or bias removal. Deterministic approaches to the problem of surface drift are often limited by strong assumptions on the underlying physics. A new approach based on linear and non-linear optimization is proposed, using hyper-ensemble deduced statistics to forecast at short time scale Lagrangian drifts from combined atmospheric and ocean operational models and local observations that were made available during the MREA04 field experiment along the West coast of Portugal. Optimization methods are based on a training/forecast cycle. The performance and the limitations of the hyper-ensembles and the individual models are discussed. Results suggest that our statistical methods reduce the position errors significantly for 12 to 48 h forecasts and hence compete with pure deterministic approaches.