(iii) Regional Climate Models (RCMs)

The GCMs above are typically run at 200-300km resolution. There is a growing set of models which work basically in the same way, but which are tuned to represent additional processes that become important when we consider the evolution of the atmosphere at smaller scales. These high resolution regional climate models (RCMs, with a grid-spacing down to 20km or less) can be driven with the forecast large scale atmospheric fields generated by the GCMs/CGCMs to potentially provide ensembles of seasonal forecasts with higher spatial resolution (e.g. see Fig. 3.4 for example of the scales of a GCM and RCM - more details will be given in lecture 4 on downscaling). This method is still largely in the research phase. It is playing a role in establishing the scientific basis for making seasonal forecasts at smaller spatial and temporal scales, and my prove to be a valuable tool for contributing to the generation of tailored forecast information for applications. These issues are returned to in detail in lecture 4.

(iv) Statistical Transformations of Model Output

The methods above (i - iii) are all numerical models based on the equations of physics and dynamics that govern the climate system. The output of such models usually have first order systematic errors which we can correct using simple statistical concepts. For example, if the long-term (30-year) mean rainfall for a particular region and season is substantially less than observed, we can correct the output by adding a constant value to the model's predicted rainfall (e.g. see Fig. 3.5). The variance can similarly be adjusted.

However, more sophisticated transformations are also possible to improve the accuracy and usefulness of the model predictions. This adjustment has long been practiced in weather forecasting, where it is described as model output statistics (MOS). Ways in which the output is transformed include the following:

i)Systematic shifts in rainfall anomalies: for any given region (e.g. East Africa) the GCM may always tend to respond to an El Niño forcing with a rainfall anomaly pattern that is shifted by a set distance from the observed anomaly. This can be corrected using regression-based techniques including coupled pattern techniques like canonical correlation analysis (CCA).

ii)The ensemble from the GCM may always underestimate the spread of possible conditions given the forecast SST - thus, the probability forecast from the GCM would need to be adjusted, based on comparison of the GCMs forecast with observations over past years.

iii)The output from a number of models can be combined into one best estimate of the likely outcome, or probability distribution of possible outcomes.

iv)The output of the model can be transformed into information that more closely matches the information needed to support decision-making in sectors like agriculture that are affected by seasonal climate variations (these aspects are further discussed in Part 4 on Downscaling). Some of this information may be for non-climate variables that are affected by the climate (e.g. inflow to a reservoir).