The challenge of monthly and seasonal forecasting consists in predicting, one or several months or seasons ahead, means of weather variables. There are two major properties of the climate system that would enable such kind of long-range predictability (Vautard et al., 1996):
- the lower boundary forcing, such as the ocean or the soil water content, that evolve on a much longer timescale than the atmosphere, being thus more predictable
- the atmospheric dynamical system itself that has regular predictable components, as persistent weather regimes or low-frequency oscillation spells.
On seasonal time-scales the prospect for atmospheric prediction is based primarily on the first premise.
In this project a comparison of the ability of dynamical multimodel forecasts is made. Also, some empirical forecasts are envisaged. The dynamical forecasts rely on both properties stated above, the SST prescription and good initial conditions. They are made by using the ensemble technique. The basic idea of ensemble forecasting is to run not just one deterministic model but to run a model many times with slightly different initial conditions. All these initial conditions lie within some phase-space region which defines the uncertainty in the analysis (Reed, 1995). This uncertainty has been introduced by initialising the model with 24-hour lagged analyses. The empirical model is based on the dynamical properties of the atmosphere because no SST nor other boundary conditions are specified.
When long-range forecasts are issued from a GCM, three sources of forecast errors are expected: errors in the initial conditions, model errors and errors in the boundary conditions (Vautard et al., 1996). Even with a perfect ocean model, the former two sources are quite important. The ECMWF reanalyses (ERA) for 1979-1993 have been used in these simulations as initial conditions, but due mainly to areas of poor data coverage, analyses are far from perfection. The dynamical models, the ARPEGE, ECMWF and UKMO GCMs, have been run for 15 forecasts of around 120 days long for each season from March 1979 to December 1993, by using prescribed SST (i.e., a perfect non-interactive ocean). To take into account the non-linear evolution of the weather simulation, the ensemble method has been used. Each forecast consists of 9 integrations lagged by 24 hours. ARPEGE is a spectral model with a T42 resolution and 31 levels. The ECMWF GCM is similar to ARPEGE but a T63 truncation and a different physical set of parameterizations have been used.
The mean spread over North America in the one-month lead seasonal range (31-120 days) is smaller than over Europe. Spread has been assessed by using anomaly correlation coefficients around the ensemble mean forecast. Skill has been computed in terms of ACC with ERA data as verification. For Europe, considered as the area 30N to 75N and 12.5W to 42.5E, the dynamical skill in the monthly and seasonal range is slightly lower than for the North American region.
There is a clear seasonal dependence of the skill. For the zero-month lead monthly range (1-30 days) the highest skill is found in spring, both for Z500 and T850, and for both models and their simple combination (multimodel combination has been made by taking the two sets of forecast anomalies from ARPEGE and ECMWF simulations, creating an 18-member ensemble). For the one-month lead seasonal range the highest skill is found for winter (Fig.1). For explaining this seasonal evolution let's consider that the extratropical atmosphere may be most predictable in summer, when internal variability is weakest. On the other hand a large component of extratropical predictability is of tropical origin, and predictable planetary-scale circulation patterns in the tropics influence extratropical circulations trough teleconnections induced by Rossby-wave dynamics which are generally largest in the winter season when meridional potential-vorticity gradients are strongest.
Next activity
Research activities
- the lower boundary forcing, such as the ocean or the soil water content, that evolve on a much longer timescale than the atmosphere, being thus more predictable
- the atmospheric dynamical system itself that has regular predictable components, as persistent weather regimes or low-frequency oscillation spells.
On seasonal time-scales the prospect for atmospheric prediction is based primarily on the first premise.
In this project a comparison of the ability of dynamical multimodel forecasts is made. Also, some empirical forecasts are envisaged. The dynamical forecasts rely on both properties stated above, the SST prescription and good initial conditions. They are made by using the ensemble technique. The basic idea of ensemble forecasting is to run not just one deterministic model but to run a model many times with slightly different initial conditions. All these initial conditions lie within some phase-space region which defines the uncertainty in the analysis (Reed, 1995). This uncertainty has been introduced by initialising the model with 24-hour lagged analyses. The empirical model is based on the dynamical properties of the atmosphere because no SST nor other boundary conditions are specified.
When long-range forecasts are issued from a GCM, three sources of forecast errors are expected: errors in the initial conditions, model errors and errors in the boundary conditions (Vautard et al., 1996). Even with a perfect ocean model, the former two sources are quite important. The ECMWF reanalyses (ERA) for 1979-1993 have been used in these simulations as initial conditions, but due mainly to areas of poor data coverage, analyses are far from perfection. The dynamical models, the ARPEGE, ECMWF and UKMO GCMs, have been run for 15 forecasts of around 120 days long for each season from March 1979 to December 1993, by using prescribed SST (i.e., a perfect non-interactive ocean). To take into account the non-linear evolution of the weather simulation, the ensemble method has been used. Each forecast consists of 9 integrations lagged by 24 hours. ARPEGE is a spectral model with a T42 resolution and 31 levels. The ECMWF GCM is similar to ARPEGE but a T63 truncation and a different physical set of parameterizations have been used.
The mean spread over North America in the one-month lead seasonal range (31-120 days) is smaller than over Europe. Spread has been assessed by using anomaly correlation coefficients around the ensemble mean forecast. Skill has been computed in terms of ACC with ERA data as verification. For Europe, considered as the area 30N to 75N and 12.5W to 42.5E, the dynamical skill in the monthly and seasonal range is slightly lower than for the North American region.
There is a clear seasonal dependence of the skill. For the zero-month lead monthly range (1-30 days) the highest skill is found in spring, both for Z500 and T850, and for both models and their simple combination (multimodel combination has been made by taking the two sets of forecast anomalies from ARPEGE and ECMWF simulations, creating an 18-member ensemble). For the one-month lead seasonal range the highest skill is found for winter (Fig.1). For explaining this seasonal evolution let's consider that the extratropical atmosphere may be most predictable in summer, when internal variability is weakest. On the other hand a large component of extratropical predictability is of tropical origin, and predictable planetary-scale circulation patterns in the tropics influence extratropical circulations trough teleconnections induced by Rossby-wave dynamics which are generally largest in the winter season when meridional potential-vorticity gradients are strongest.