A novel way of correcting between-ensemble-member biases in a lagged S2S ensemble

Abstract

<p>The UK Met Office seasonal forecast system, Global Seasonal Forecast System version 5 (GloSea5), is an ensemble forecast prediction system providing sub-seasonal and seasonal forecasts over the globe with ~60 km resolution in the mid-latitudes. GloSea5 also produces hindcasts or historical re-forecasts. The system produces 4 members a day, initialised at 00UTC. Two members run out to 64 days and two run out to 216 days. We use these four members to generate a 40-member lagged ensemble with 10 days of lag time, i.e. for any forecast horizon the oldest members are always 10 days older. Due to this lag and the way these ensemble members are initialised, there is a considerable within-ensemble bias, even for a nominal “day 1” forecast. This within-ensemble bias evolves with increasing lead time horizon.</p><p>Traditionally hindcasts are used to correct for the so-called model drift. In this work the idea of using a distribution of daily rainfall amounts from short-lead time forecasts is used using the 2019 Indian monsoon season. Quantile mapping is trialled as a means of removing the “within-ensemble-member” bias to ensure that all ensemble members are drawn from a more consistent underlying distribution. Achieving this would enable the members to be used to drive downstream applications such as hazard or impact models, as such models require individual ensemble members.</p><p>The presentation will demonstrate the methodology and the impact it has on ensemble forecast skill, complementing the presentation by Kolusu et al. (same session in conference) which presents an evaluation methodology focusing on patterns for different accumulation lengths and forecast horizons.</p>