UST's AI forecast model predicts rainstorms up to four hours in advance

The Hong Kong University of Science and Technology on Wednesday said its researchers have developed an AI model to forecast dangerous convective storms.

Adverse weather has become a global issue that countries hope to tackle by minimising its impacts on societies.

A HKUST research team has developed the world's first AI model, known as Deep Diffusion Model of Satellite Data, to better forecast rapidly-evolving weather such as rainstorms, thunderstorms and tornadoes.

Compared to the conventional models, which rely on numerical prediction methods, the AI system uses satellite data that monitors cloud evolution from space, allowing the model to forecast storm development with a lead time of four hours, instead of the present models giving only 20 minutes to two hours warning time.

The system also improves the forecast's accuracy rate by more than 15 percent.

The deep diffusion is a relatively newcomer in the AI techniques," said Su Hui, chair professor of HKUST Department of Civil and Environmental Engineering. "We use that to represent this kind of chaotic and randomness in the weather forecast."

While the team has trained the model through data collected by China's FengYun-4A satellite, Su mentioned observation using different channels as the key aspect of improving the system's accuracy and lead time.

"Observations from satellite, from ground based instrument, and also potentially with aircraft measurement from a lot of UAVs," she said. "So if we have a low-altitude economy, we'll have a lot of drones flying. We can also put sensors on those UAVs, so we can collect more data."

The team has reached out to the Observatory and shared their AI model in the hope of collaborating with the institution.