Tim Palmer

Abstract:

Increasing concentrations of atmospheric carbon dioxide will almost certainly lead to changes in global mean climate. But because--by definition--extreme events are rare, it is significantly more difficult to quantify the risk of extremes. Ensemble-based probabilistic predictions, as used in short- and medium-term forecasts of weather and climate, are more useful than deterministic forecasts using a 'best guess' scenario to address this sort of problem. Here we present a probabilistic analysis of 19 global climate model simulations with a generic binary decision model. We estimate that the probability of total boreal winter precipitation exceeding two standard deviations above normal will increase by a factor of five over parts of the UK over the next 100 years. We find similar increases in probability for the Asian monsoon region in boreal summer, with implications for flooding in Bangladesh. Further practical applications of our techniques would be helped by the use of larger ensembles (for a more complete sampling of model uncertainty) and a wider range of scenarios at a resolution adequate to analyse average-size river basins.

Abstract:

Operational forecasting is hampered both by the rapid divergence of nearby initial conditions and by error in the underlying model. Interest in chaos has fuelled much work on the first of these two issues; this paper focuses on the second. A new approach to quantifying state-dependent model error, the local model drift, is derived and deployed both in examples and in operational numerical weather prediction models. A simple law is derived to relate model error to likely shadowing performance (how long the model can stay close to the observations). Imperfect model experiments are used to contrast the performance of truncated models relative to a high resolution run, and the operational model relative to the analysis. In both cases the component of forecast error due to state-dependent model error tends to grow as the square-root of forecast time, and provides a major source of error out to three days. These initial results suggest that model error plays a major role and calls for further research in quantifying both the local model drift and expected shadowing times.