Multi-analysis forecasts

To investigate the effect of analysis differences on medium-range forecasts, the ECMWF model is run routinely from initial conditions of other operational forecast centres.

• Aim
• Independent analyses
• Construction of initial conditions
• Preliminary results

Aim of the project

A key aim of this project is to investigate the sensitivity of the forecast to analysis differences. During the next year, ECMWF will conduct an investigation of the performance of the system, including a comparison with the operational EPS.

Forecasts from independent analyses

ECMWF is now routinely running forecasts using the ECMWF model initialised with the 12 UTC analyses from Deutsche Wetterdienst (DWD), Météo-France, the National Centers for Environmental Prediction (NCEP) and the Met Office (UK). A 10-day forecast is run from each analysis and from a 'consensus' analysis generated as a simple average over all the available analyses, including the ECMWF analysis. The forecasts are run at TL255L40 resolution using the same configuration of the ECMWF model as is used for the EPS. This set of forecasts is referred to as the multi-analysis (MA) system.

Specification of initial conditions

Because the analyses generated by different centres are not directly compatible (for example they use different horizontal grid spacing, different vertical levels and different orography), interpolation is needed to produce an analysis on the ECMWF model grid. This is done using the following procedure. Data is received from each centre as fields of wind, temperature and humidity on standard pressure levels. The difference between this pressure level analysis and the ECMWF pressure level analysis is calculated and then interpolated onto the ECMWF model levels. The resulting perturbation is added to the ECMWF analysis to produce an approximation to the analysis of the other centre. This method has been chosen to minimise the disturbance to the analysis fields - any inconsistencies are in the small perturbations rather than in the full fields. Only the upper-air fields are perturbed; there is no perturbation to the ECMWF surface fields. The only exception to the above procedure is for the Météo-France analysis, which is provided directly on the ECMWF grid.

Preliminary results

The following figures show first results from a preliminary evaluation of the performance of the multi-analysis system. Scores are calculated for 500 hPa height fields for 57 cases from July and August 2001 (some analysis data was missing for the remaining 5 dates).

Figure 1 shows the average anomaly correlation over the Northern Hemisphere for each member of the MA system and for some of the corresponding operational forecast from the centre providing the analysis. There are clear differences in overall performance between the MA forecasts started from different analyses (Fig 1d). Over this set of cases the predictions from the DWD and Météo-France analyses have the lowest scores while the NCEP analyses provides the most skilful forecast. The performance of the MA-forecast from the DWD analysis is similar to that of the DWD operational forecast (Fig 1a; the DWD forecast extends to 7 days ahead) and the same correspondence can be seen for the forecasts from the NCEP (Fig 1b) and UK (Fig 1c) analyses (the Météo-France operational forecast is made only for 3 days ahead and is not shown).

Figure 1.

Figure 2 compares the day-to-day variability in performance of each MA member with the corresponding operational forecast over Europe at day 3. Here again the similarity between forecasts from the same analysis (but using different models) is clear.

Figure 2

By day 6 (Fig 3) there are cases where there are differences in skill between the MA-member and the operational forecast from the same analysis, but such examples are relatively infrequent.

Figure 3