In August 2023, Tropical Storm Hilary made landfall in Mexico and Southern California, prompting Southern California’s first ever Tropical Storm Watch. Even though the storm weakened from its peak Category 4 Hurricane status, it still brought devastating rainfall and flooding to a part of the world that rarely sees tropical rainfall, especially in summer.
Accurate information about how such storms evolve can play an outsized role in helping businesses and societies prepare for their impact, and we took a look at how well our AI models did in forecasting the storm.
Excarta’s AI models accurately predicted Tropical Storm Hilary’s track and intensity 6 days before it made landfall, demonstrating the value of AI models in providing actionable weather intelligence for extreme events.
What did we examine?
In particular, we took a quick look at the following aspects of the storm:
- Location: How well did our models capture the track of the storm? Did our models do a good job predicting where it would make landfall?
- Intensity: How well did our models predict the intensity of the storm?
- Precipitation: Since excessive rainfall and flooding were the cause of much of the damage, did our models capture the intensity, timing, and location of the rainfall?
Hurricanes and tropical storms impact a large area, and the more advance notice we get about them, the better prepared we can be. So, we specifically checked how well our models were able to forecast the storm 5 days before it actually made landfall. Since the storm made landfall around August 20th, 18:00 UTC, we chose the forecast issued by our models at August 15th 00:00 UTC to investigate. For comparison against a baseline, we chose the forecast issued by the NOAA’s GFS model at the same time.
What actually happened
To see what actually happened, we use the data from ECMWF’s ERA5 reanalysis dataset1. The ERA5 dataset combines ground truth data from several sources with a state-of-the-art physics model to provide a comprehensive state of the atmosphere for historical weather.
The image below is a visualization of sea-level pressure, with the purple areas indicating regions of low pressure, and orange areas indicating regions of high pressure. The cyan line indicates the track of the storm, and the eye of the storm is the area with the lowest pressure, colored in the deepest purple. One can see the telltale “eye” with the lowest pressure moving along the track, and the storm gradually weakening as it nears landfall.
Similar to the image above, we can see a similar visualization for wind speeds. Here, the deeper blues and purples indicate a lower wind speed, and the brighter oranges indicate higher wind speeds. The highest wind speeds are seen near the eye of the storm, and gradually weaken as the storm nears landfall.
What Excarta’s forecasts said
We can now look at what Excarta’s models and NOAA’s GFS model predicted for the storm. The forecasts from both Excarta and GFS were issued at August 15, 00:00 UTC, roughly 6 days before the storm made landfall. A dark blue dot is used to indicate the predicted eye of the storm in both the Excarta and GFS forecasts.
The images below compare the predicted mean sea-level pressure from Excarta against the ground truth (ERA5, right column). As can be seen, Excarta’s model does remarkably well in predicting how the storm evolves, predicting with high accuracy both the track of the storm itself, as well as when it is expected to make landfall. It also captures the weakening intensity of the storm.
The image below compares the mean sea-level pressure forecast from GFS (left column) with the same ground truth as above. The GFS forecast expects the storm to stay very intense, and move less slowly, effectively predicting a later and potentially strongter landfall.
When comparing at wind speeds to the ground truth as well, we see that Excarta’s models do really well in capturing where the winds are expected to be high, and when they will dissipate.
Here as well, the GFS forecast overestimates the wind speeds and expects the storm to move more slowly than it actually did.
Lastly, we take a look at the predicted rainfall intensity and patterns. Once again, Excarta’s models are able to capture both the intensity and patterns of the expected rainfall from the storm, almost a week ahead of time.
Extreme events, by definition, happen rarely — but their impact makes predicting them accurately especially important. The fact that they happen relatively rarely is a blessing, but also makes them a potential challenge for AI models.
Tropical Storm Hilary was a rare event not just because it was a hurricane, but because it maintained a great deal of its strength in a part of the world that doesn’t often get such storms. Seeing how closely Excarta’s models were able to predict the storm’s track is a good demonstration of the ability of AI models to predict weather — even extreme weather — with great accuracy.