High Resolution Forecasts For The Whole World
We recently improved our forecasting model to produce global high-resolution forecasts for key variables. In this post, we'll explore how we evaluate the quality of high-resolution forecasts, and how we compare against competing high-quality forecasts.
- Compared to GFS forecasts at 0.25 deg resolution, and HRES forecasts at 0.1 deg resolution, Excarta's High-Res forecasts are at a 0.025 deg resolution. This provides unparalleled detailed information not available in other global forecasting models.
- We compare our high-res model with the state of the art high-res model from NOAA, and see significant improvements in forecasting accuracy.
Why are high-res weather forecasts important?
Weather forecasting models use a grid to divide the world into cells of a fixed size, with each grid cell corresponding to a specific part of the Earth. Large grid cells cover larger areas (e.g. all of Manhattan), which makes it hard to capture weather patterns inside a grid cell. Smaller grid cells cover smaller parts of the Earth (e.g., just downtown Manhattan) leading to fewer approximations and more accurate forecasts. A high-resolution forecast is thus a crucial part of delivering accurate weather forecasts.
Global weather models like GFS are typically run with grid sizes of approximately 25km, which creates relatively coarse forecasts, illustrated below:
In contrast, a high-res forecast from Excarta for the same area would look smoother, capturing finer variations:
The challenge for high-res weather forecasts.
Making the grid cells smaller increases the resolution and accuracy of forecasts, but also increases the cost of producing these forecasts. As a result, conventional high-res models are typically only available for specific regions (e.g. continental US, or parts of Europe), resulting in a shortage of high-res forecasts for many parts of the world. For the continental US (CONUS), the best publicly available high-res forecast comes from NOAA's HRRR model, which provides hourly forecasts for up to 48 hours, at a 3km resolution.
On the other hand, Excarta's weather models are able to provide high-res forecasts for the entire globe by harnessing the power of AI. Producing forecasts at 0.025 degree resolution (roughly 2.5km), Excarta's forecasts provide access to high-quality weather intelligence for businesses worldwide, overcoming the limitations of expensive regional model forecasts.
Evaluating the quality of our forecasts
To evaluate the quality of our high-res forecasts, we conducted exhaustive comparisons between our forecasts and on-the-ground observations. NOAA's Meteorological Assimilation Data Ingest System or MADIS program  collects weather data from over 100,000 stations around the world, giving us a rich source of observations to test against. The snapshot below, showing MADIS weather stations in the northeast US, illustrates the density of the observations used in our evaluations.
We take care to only use MADIS observations that pass rigorous quality checks, comparing HRRR forecasts to Excarta's forecasts for the continental US. These evaluations are performed for multiple months to discount for any effect of seasonality on model predictions.
While Excarta's AI-powered weather products are capable of predicting many different variables, we focus here on our evaluations for wind speeds. We measure the accuracy of forecasts by calculating the Root Mean Squared Error (RMSE) of the predicted and actual wind speeds. Since RMSE measures error, an RMSE of 0 indicates a perfect forecast.
Comparing the RMSE of HRRR vs. Excarta shows that Excarta's forecasts have a significantly lower error over the entire 48 hour forecasting period, with as much as 25% lower over longer lead times.
Since numerical high-resolution weather forecasts are expensive to compute and not available for many parts of the world, our High-Res model provide businesses a significant advantage, especially in data-poor regions such as EMEA, South Asia and APAC.
If you're interested in learning more or speaking with us, please schedule a chat or email us at firstname.lastname@example.org.