Using Modified Bred Vector Ensembles to Improve Forecasting in Multi-Scale Dynamical Systems

Improving forecast predictability in chaotic multi-scale systems over large timescales is generally an extremely difficult problem in ensemble forecasting. However, it is generally the small scale, rapid instabilities that cascade and drive changes in the large-scale flow. Consequently, modelling and accounting for a variety of small uncertainties is critical for developing reliable long-term ensemble forecasts. We investigate the performance of Bred Vectors for generating ensemble perturbations in the multi-scale Lorenz-96 system and how the dynamics of the fast sub-system impact predictability over the slow, long term system. Consequently, we propose modifications to the Bred Vectors that sample the fast subspace more efficiently but in a dynamically informed way to significantly improve both ensemble spread and reliability while reducing error in the forecast.