Optimizing weather forecasts using models blending: a comparative analysis of Frogcast’s performances

null

null

null

It is well known that the chaotic nature of the atmosphere makes forecasting its evolution tricky. Using multiple sources of information can help tackle the issue by quantifying and reducing the uncertainty. But, how does Frogcast perform?

In this article, we present an assessment of Frogcast predictions versus individual numerical weather prediction (NWP) models outputs for 6 atmospheric variables over a 3-month evaluation period.

Formula for calculating the MAE (Mean Absolute Error)

Data and evaluation framework

General evaluation features

In order to get relevant and robust results, the following evaluation is done over a 3-month period, ranging from October 1st to December 31st 2023. The geographical domain covers most of Europe (excluding Scandinavia) and part of Magreb. It extends from 30°N to 60°N in latitude and from 15°W to 30°E in longitude.

We selected 6 classical weather variables to get a clear and wide picture of Frogcast’s performances: 2m temperature (°C), 10m wind speed (m/s), 1h precipitation rate (mm/h), 2m relative humidity (%), mean sea level pressure (hPa) and global horizontal irradiation (W/m²).

The evaluation criterion is a simple mean absolute error computed using the hourly time series.

null

null

Gridded reference data

ERA5 ECMWF reanalysis data has been used as gridded pseudo observations to evaluate the different forecasts. This dataset provides information at hourly time step on a 0.25° spatial grid. It combines model data with observations to generate a new best estimate of the state of the atmosphere. More information can be found on the Copernicus Climate Data Store website.

null

null

Frogcast vs NWP models

Frogcast’s outputs are evaluated against a set of 7 individual Numerical Weather Prediction (NWP) models: ARPEGE-EU and AROME (Météo-France), ICON-EU and ICON-D2 (DWD), GFS (NCEP), GDPS (CMC) and IFS-HRES (ECMWF). For each of them, forecasts are extracted once a day (0 UTC run) for the first 24 hourly lead times. In order to make a fair comparison, all models are reprojected and upscaled to the ERA5 0.25° grid.

null

Analyzing spatial performances across variables

null

Spatial performances for 2m temperature are presented on Figure 1 in both absolute value (a) and absolute difference between individual models and Frogcast (b). For this variable, the best scores are achieved by ICON, ICON-D2 and IFS models. However, for every grid point of the domain, Frogcast outperforms all individual models and significantly reduces the MAE, especially over continental areas.

null

null

null

null

Figure 1: MAE temperature maps for each individual model in absolute value and relatively to Frogcast.

null

null

The results regarding all 6 variables are gathered on tables 1 to 4. Scores related to ARPEGE, ICON, GFS and GDPS are averaged over the same full domain and can be compared (Table 1). For the three remaining models, scores correspond to their specific regional domain and must be considered separately.

For all atmospheric parameters, Frogcast always outperforms every single NWP model. Improvements are significant and range, for instance, from 0.2 to 0.5°C (+25% to +40%) for 2 m temperature or from 0.29 to 0.41 m/s (+24% to +39%) for 10 m wind speed. These differences vary a lot spatially with best performances of Frogcast over continental areas (not shown).

null

null

null

null

Table 1: MAE scores on full domain

null

null

null

Table 2: MAE scores on AROME’s domain

null

null

null

Table 3: MAE scores on ICON-D2’s domain

null

null

null

Table 4: MAE scores on IFS’s domain

Frogcast leads the way in accurate weather forecasting

null

Frogcast provides high-performance probabilistic weather forecasts for any location in the world and a wide range of atmospheric parameters. This study highlights the relevance of its outputs and that it outperforms every individual NWP model. Frogcast’s algorithms are optimized for every location and weather variable assuring the best forecasts for any of your specific needs.

null