Global seasonal forecasting system (GloSea5)

Ensemble forecasting method

Each forecast starts from the observed state of the ocean, land and atmosphere - the 'initial conditions'. Unfortunately, the initial conditions are not known precisely (one reason is that observing stations/instruments are sparsely distributed in some regions), and individual forecasts can be very sensitive to small uncertainties in the initial conditions. The sensitivity to initial conditions is taken into account by making a number of predictions each with different starting conditions which reflect our uncertainty in the initial state. Like its predecessor, GloSea5 uses the lagged-start ensemble generation technique (MacLachlan et al) to account for initial-condition uncertainty. 

In addition to forecast sensitivity to starting conditions, forecasts can also be sensitive to the way we represent the real climate system in the computer models - so-called sensitivities to model formulation. We address these uncertainties by using a random parameter perturbation scheme (Bowler et al, 2008) to simulate, using a single model framework, a range of representations of physical processes, thus sampling a large part of model uncertainty.

Such techniques generate an 'ensemble' of forecasts, with each individual forecast referred to as an ensemble 'member'. The forecast products are created by analysing the output from all the ensemble members.

C. MacLachlan et al, in preparation: GloSea5: the Met Office high resolution seasonal forecast

Bowler, N.E., Arribas, A., Mylne, K.R., Robertson, K.B. and Beare, S.E., 2008: The MOGREPS short-range ensemble prediction system. QJR Meteorol Soc, 134, 703-722.

Met Office GloSea5 system

The forecast products displayed are from the Met Office global seasonal prediction system, version 5, referred to as 'GloSea5'. An extensive set of performance validation information is provided for the GloSea5 system from analysis of retrospective forecasts over the period 1996-2009, updated each month.

The products are based on the output from forecasts made using a coupled ocean-atmosphere General Circulation Model (GCM). This is is a version of the HadGEM3 climate model, which uses the UM (Met Office Unified Model) atmosphere, NEMO (Nucleus for European Modelling of the Ocean) ocean and MOSES (Met Office Surface Exchange Scheme) land surface. The model resolution, as used in GloSea5, is N216 (approx 0.8° in latitude and 0.5° in longitude) - approx 50 km in mid-latitudes, in the horizontal - and 85 levels in the vertical for the atmosphere, and the ORCA0.25 grid (0.25°) and 75 levels in the vertical for the ocean. The model contains no flux corrections or relaxations to climatology. Climate forcings (aerosols, methane and CO2 concentrations, etc.) are set to observed values for the hindcast period and follow the scenario A1B afterwards. Ozone is fixed to observed climatological values and includes a seasonal cycle.

Each forecast requires initial ocean, land and atmosphere conditions. The land and atmosphere conditions are specified from atmospheric analyses that are produced separately for weather prediction purposes. The ocean initial conditions are taken from ocean analyses generated specifically for seasonal forecasting, using the 3-dimensional variational ocean data assimilation based on the multi-institution NEMOVAR project (Mogensen et al, 2009 describes a similar implementation of this method) .

Each day forecasts are started from observed conditions corresponding to that day, to create a 2-member 'perturbed physics' ensemble: the different ensemble members are created by randomly perturbing parameters in physical parametrisations. By pooling three weeks' perturbed-physics ensembles together, a larger, lagged-start ensemble is created, which samples both initial-condition and model uncertainties.

In parallel to the forecast, four times a month (on the 1st, 9th, 17th and 25th) a three-member hindcast is generated in the same way, for the 14 years 1996-2009. The hindcasts are used to calibrate the ensemble forecast, which is then used to generate the seasonal products.

K.S. Mogensen, M.A. Balmaseda, A. Weaver, M.J. Martin, A. Vidard. NEMOVAR: A variational data assimilation system for the NEMO ocean model. ECMWF newsletter, summer 2009.

Last updated: 16 August 2013