Review of Typhoon No. 30 Track Forecasts: AI Models Outperform Traditional Models

AI models tracked the actual path closely from an early stage

Each thin line in the chart accompanying this article represents the result of a numerical simulation run by a physical or AI model from a meteorological agency or other organization around the world. The lines connect the low-pressure system’s center positions extracted from ensemble forecast simulations. By running many simulations with slightly different initial conditions, forecasters can use the results to assess forecast confidence. Traditional physical weather models predict future weather by using computers to solve equations based on meteorological theory. AI weather models, by contrast, learn how the atmosphere changes over time from past atmospheric data used as training targets, then use the trained models to predict future weather conditions.

A comparison shows that the traditional physical models initially tended to forecast a track that turned northeast from south of Japan and headed east of the country. This trend changed sharply from around the 16th, with more simulations directing the storm from near Taiwan toward the South China Sea. Only after the 20th did forecasts for a northward track across the East China Sea become more common.

By contrast, simulations using AI technology showed little change in their overall forecast trend from the outset, making their predictions more consistent. They continued to forecast a northward track across the East China Sea from around the 14th, before the typhoon formed.

Ultimately, AI models substantially outperformed traditional physical models in forecasting the track of Typhoon No. 30. Weathernews’ Global Storm Center conducts quantitative evaluations of typhoon forecast models. It will publish evaluations of storms other than No. 30 once they have been compiled.