Using deep learning to better forecast tropical cyclone tracks and intensities

Tropical cyclones are among the deadliest natural disasters, and the most expensive to recover from. Accurately predicting their tracks and intensities can improve early warning systems, thereby reducing casualties and damage.

Recent approaches using deep learning (DL) for weather forecasting struggle to predict tropical cyclone intensity. Zeng et al. integrated DL with a traditional, physics-based Weather Research and Forecasting (WRF) model to determine whether augmenting physics-based simulations with DL can improve cyclone track and intensity predictions.

In their tests comparing five DL-integrated WRF (DLWRF) models with two conventional WRF models, they found the hybrid approach maintains consistently stable performance regardless of the DL model used. Moreover, the DLWRF techniques consistently outperformed WRF in track prediction, and demonstrated an improvement in intensity prediction compared to pure DL-based models, comparable to WRF.

“The hybrid approach capitalizes on the respective advantages of each method,” said author Zheng-Wei Chen.

These improvements make DLWRF a useful alternative or supplement to traditional WRF methods for tropical cyclone forecasting. The researchers — who operate an urban meteorological monitoring network in Hong Kong as part of their work — plan to continue exploring how DL can further optimize forecasting accuracy for tropical cyclones and establish an enhanced early-warning system, crucial for the long-term sustainability of urban coastal areas.

“These findings may provide a practical application strategy,” said author Yi-Qing Ni. “We can use cost-effective pure DL models for an initial track assessment and then employ the deep learning-integrated WRF model to generate more reliable, high-resolution intensity forecasts.”

Source: “Evaluating deep learning-integrated physics-based models for tropical cyclone track and intensity predictions,” by Yuan-Jiang Zeng, Hao-Yan Liu, Guang-Zhi Zeng, Yi-Qing Ni, Zheng-Wei Chen, and Pak-Wai Chan, Physics of Fluids (2026). The article can be accessed at https://doi.org/10.1063/5.0303579 .