What can a decadal prediction system predict? From a climate simulation to climate information


In order to potentially predict a wanted climate information on the decadal time-scale, then this climate information must be in the memory of the different components of the climate system. Furthermore, the initial conditions and the basic physical laws must be well known. Therefore, predictability varies for different climate variables, time- and spatial scales. In certain cases, no predictability can be found.

The global earth system model outputs results for all simulated variables of the climate system, such as precipitation, ocean temperatures or soil moisture. These are calculated for all grid boxes of a pre-defined resolution for the globe, e.g., 200km x 200km on the earth surface and in the different height levels in the atmosphere and in the ocean. The different variables can be output and saved according to needs on a 3-hourly, 6-hourly, daily, weekly or monthly time resolution.

The most common variables to analyse are those that are known from a typical weather report: temperature, precipitation and wind. Since these variables have been observed for several decades, the predictions of these can be properly evaluated. Many other variables are interesting from a scientific perspective, especially in the ocean, since these in theory show important memory for decadal predictability.

For the variables temperature, precipitation and wind, one can already from theoretical considerations learn something about the different levels of predictability. Since air temperature is in general more homogeneous over large spatial and temporal scales, it is in theory more predictable than precipitation – a variable that can be strongly variable over small spatial and temporal scales. The predictability of wind variables is dependent on whether one consider rather large-scale homogeneous wind fields or rather small-scale variable wind fields, the former being more predictable.

Furthermore, the predictability of a given variable also depends on which region and which time period is considered.

The aim of a decadal prediction system is to use the memory of those components of the climate system that act on a time-scale from years up to a decade and gain predictability from this memory. Predictability can vary from year to year, when the interplay of atmospheric processes changes due to the large-scale situation. To make statistically sound statements, the climate predictions are usually made for averages of aggregated time periods and the future predictability is based on the analysis of long past time periods.

The influence of the observed initial conditions is very strong in the first forecast years, but is reduced with increasing forecast lead time, so that the predictability for the later forecast periods often only results from the influence of the boundary conditions. The predictability is thus generally reduced with increasing forecast lead time. In the model world this is not always so clear due to erroneous initial conditions, insufficient process understanding or errors in the numerical calculations. 

In MiKlip the standard approach is to consider averages over the lead times year 1-4, year 2-5, …, year 7-10. This allows sufficiently long time periods for the averaging and allows the analysis of “near” and “distant” predictions. The first few years can show a strong influence from the initialisation and as such the first and second half of the decade can show different prediction skill.

Climate processes in different regions of the world are driven by feedbacks between different components of the climate system, therefore the predictability of a given climate variable may also depend on the chosen region. For instance, a range of evaluation methods show that sea surface temperatures in the North Atlantic exhibit higher forecast skill than other regions. This insight is especially valuable, since North Atlantic sea surface temperatures have a large influence on European climate. Temperatures over Europe generally show a trend, which is connected to a climate change signal. This trend is to a large extent captured by the MiKlip system and the climate projections.

As is the case for the time scale, the different memories of the components of the climate system mean that decadal predictions systems are suited rather for the consideration of larger spatial scales. On smaller spatial scales local variability that is only weakly connected to the larger scale phenomena might occur. Noise from smaller scales can further reduce the predictability. Therefore, to increase the statistical significance of the climate predictions these are usually considered over larger spatial scales, e.g. over 250-500km, the predictions are thus aggregated over several grid boxes. The spatial scale of the aggregation in turn depends on the chosen variable.   


There is a need to aggregate values over larger regions. Here illustrated for the North Atlantic.

In a first instance, the decadal predictions are calculated with a global climate model with coarse spatial resolution (ca. 200km x 200km). If one wants to analyse the predictions on smaller regional scales, two methods can be used. With  dynamical downscaling the data from the global model are brought to a higher resolution via a regional climate model. The regional models are constructed according to the same physical principles and basic equations, however they are used with a higher resolution (25-50km) and due to computing restrictions only over a smaller region (e.g., Europe). With statistical downscaling the increase in resolution of the data from the global model is obtained by using statistical relationships between large-scale predictors (influencing variables), such as air pressure, and small-scale target variables, such as precipitation. These statistical relationships are obtained using observational data. 

The dependency of the predictability on variables, spatial and time-scales described above still hold here. One must test, whether predictability can be found on this more resolved spatial scale and how the predictability changes in relation to the coarser spatial resolution.

More information on this topic can be found in the section on research on regionalisation.

In order to get from the raw model output to the wanted climate information it is thus necessary to make several decisions: The choice of variable, a region and a time-period. The following example shows the prediction of the global MiKlip climate prediction system for the temperature above the North Atlantic (spatial resolution 200km) for 4-year means. This example of a prediction will be evaluated in terms of its prediction quality in the following section. For a full explanation of this graphic, visit the forecasts webpage.

This graphic illustrates a forecast of ensemble mean surface temperature anomalies for the North Atlantic put in relation to past observations. A black and white scale is used here to indicate that this is a raw forecast without any information on the evaluation of the system. Before using and issuing forecasts from a decadal climate prediction system, the system needs to be carefully evaluated against past observations.