8.4.11 Shorter-Term Predictions Using Climate Models
This subsection focuses on the few results of initial value predictions made using models that are identical, or very close to, the models used in other chapters of this report for understanding and predicting climate change.
Since the TAR, it has been shown that climate models can be integrated as weather prediction models if they are initialised appropriately (Phillips et al., 2004). This advance appears to be due to: (i) improvements in the forecast model analyses and (ii) increases in the climate model spatial resolution. An advantage of testing a model’s ability to predict weather is that some of the sub-grid scale physical processes that are parametrized in models (e.g., cloud formation, convection) can be evaluated on time scales characteristic of those processes, without the complication of feedbacks from these processes altering the underlying state of the atmosphere (Pope and Stratton, 2002; Boyle et al., 2005; Williamson et al., 2005; Martin et al., 2006). Full use can be made of the plentiful meteorological data sets and observations from specialised field experiments. According to these studies, some of the biases found in climate simulations are also evident in the analysis of their weather forecasts. This suggests that ongoing improvements in model formulation driven primarily by the needs of weather forecasting may lead also to more reliable climate predictions.
Verification of seasonal-range predictions provides a direct test of a model’s ability to represent the physical and dynamical processes controlling (unforced) fluctuations in the climate system. Satisfactory prediction of variations in key climate signals such as ENSO and its global teleconnections provides evidence that such features are realistically represented in long-term forced climate simulations.
A version of the HadCM3 AOGCM (known as GloSea) has been assessed for skill in predicting observed seasonal climate variations (Davey et al., 2002; Graham et al., 2005). Graham et al. (2005) analysed 43 years of retrospective six-month forecasts (‘hindcasts’) with GloSea, run from observed ocean-land-atmosphere initial conditions. A nine-member ensemble was used to sample uncertainty in the initial conditions. Conclusions relevant to HadCM3 include: (i) the model is able to reproduce observed large-scale lagged responses to ENSO events in the tropical Atlantic and Indian Ocean SSTs; and (ii) the model can realistically predict anomaly patterns in North Atlantic SSTs, shown to have important links with the NAO and seasonal temperature anomalies over Europe.
The GFDL-CM2.0 AOGCM has also been assessed for seasonal prediction. Twelve-month retrospective and contemporaneous forecasts were produced using a six-member ensemble over 15 years starting in 1991. The forecasts were initialised using global ocean data assimilation (Derber and Rosati, 1989; Rosati et al., 1997) and observed atmospheric forcing, combined with atmospheric initial conditions derived from the atmospheric component of the model forced with observed SSTs. Results indicated considerable model skill out to 12 months for ENSO prediction (see http://www.gfdl.noaa.gov/~rgg/si_workdir/Forecasts.html). Global teleconnections, as diagnosed from the NCEP reanalysis (GFDL GAMDT, 2004), were evident throughout the 12-month forecasts.