This website uses cookies. Read about how we use cookies.

Close window

Increased resolution global forecast models

An example of a Met Office global probabilistic ensemble forecast over parts of Australia and Indonesia, plotted at model resolution (July 2017).

September 2017 – As a result of the recent significant investment in our supercomputer capacity, we have been able to introduce a major upgrade to our global Numerical Weather Prediction (NWP) models, increasing the horizontal resolution of both the deterministic and ensemble forecasts.

This is driven by the overarching strategy for delivering global to local information on the likelihood and characteristics of hazardous weather and its translation to impact, for all forecast lead-times from days to decades. The new high resolution model leads to improved global forecast skill and has also been shown to improve our regional models, such as our 1.5 km resolution UK Model, which are nested within the global model.

Model changes

The global deterministic forecast model’s grid-spacing has reduced from 17 km to 10 km (Figure 1). This allows us to more accurately represent coastlines, topographic features and provide more structural detail to developing weather systems. Some products will be derived from a 17 km re-gridded forecast dataset, but the benefits of the higher model resolution will still be evident in these forecast products.

Figure 1: Global model land-sea mask and orography for the 17km model (left) and 10km model (right).

We have also improved the global ensemble model’s grid-spacing from 33 km to 20 km and have increased the number of ensemble members generated every 6 hours from 12 to 18, allowing us to produce ensemble forecasts with 36 rather than 24 members. This provides more accurate estimates of forecast uncertainty and probabilities of significant weather events, and an improved range of scenarios for the intensity and track of individual weather systems.

Data assimilation changes

We have made some changes to the use of satellite data in our NWP system, which includes adding more observations from new instruments as well as from additional channels on existing instruments. These all contribute incrementally to an improved atmospheric analysis and in turn to an improved forecast. We now assimilate ground-based Global Navigation Satellite System (GNSS)  observations over the USA and have improved bias-correction and observation error estimates for these observations. GNSS signals are used to derive information on the temperature and humidity structure of the atmosphere. We also now use a subset of infrared sounder radiance  channels over high land from the Atmospheric Infrared Sounder (AIRS), Cross-track Infrared Sounder (CrIS) and Infrared Atmospheric Sounding Interferometer (IASI) instruments.. Other changes include correcting an issue in the use of the snow analysis in the forecast model and the upgrading of the radiative transfer model  to version RTTOV-11. This model is required to carry out rapid radiative transfer calculations in order to assimilate satellite radiance data in NWP models.

Forecast improvements

The new global forecasting system was tested in our Parallel Suite 39 (PS39) during the period April to June 2017, and was compared to the operational forecasting system. The change results in improved global forecast skill for winds, heights and pressures as measured against both observations and analyses, which contributes to the steady incremental improvement to our global forecasts. There are larger improvements to near-surface weather parameters, particularly winds and temperatures, which have seen improvements of around 1% for UK stations and about 4% for tropical stations. During the final package tests for the winter season (January to March) 2016 we saw significant improvements to many forecast parameters with reduced bias and root-mean-square-error, e.g. 850 hPa temperatures (Figure 2). The top-left panel in this plot shows that the negative temperature bias in OS38 is reduced (closer to zero) in PS39 and that the root-mean-square error (bottom-left panel) is reduced at all forecast lead-times. The panels on the right show the difference from PS39 against OS38 and the error-bars indicate the level of statistical significance of these differences.


Figure 2: Mean (top row) and root-mean-square error (bottom row) of 850hPa temperature forecasts against analyses versus lead-time for the control (red) and PS39 test (blue) during January to March 2016. Relative change against the control is shown in the right-hand column.

The increased number of ensemble members improves the reliability of probabilistic forecasts. This is seen by an improvement of up to 3% in the continuous ranked probability score (CRPS) for near-surface forecast fields in the PS39 global ensemble. CRPS measures how good the ensemble probability forecasts are (a lower score is better in this metric). Higher resolution forecasts (blue) improve probabilistic forecast skill against current system (red) by over a day lead-time for near-surface wind, and having more ensemble members (green) gives an additional ½ day improvement (Figure 3). From the ensemble improvements, we also expect improved guidance for alternate forecast scenarios, such as forecast tracks of tropical cyclones. In the example in Figure 4, the rather unusual cyclone track is barely captured in the smaller OS38 ensemble (red) but is well-represented, albeit as a lower probability scenario, in the PS39 ensemble (blue).

Figure 3: CRPS with forecast range of 10m wind forecasts against station observations in the Northern Hemisphere for the 33 km 12-member operational ensemble (red), 20 km 12-member ensemble (blue) and 20 km 18-member ensemble (green) for the PS39 testing period April to June 2017.


Figure 4: Individual ensemble member tropical cyclone track forecasts (left) and storm strike probability (right) for the operational global ensemble system (top row) and PS39 (bottom row).





Last updated:

Follow us on

Facebook, Twitter, YouTube, Instagram, Snapchat, or LinkedIn Facebook Follow @metoffice on Twitter YouTube Instagram Snapchat LinkedIn