USP’s TITAN-LSTM Tracks Storms 30 Minutes Ahead on a Laptop

Forecasting exactly where a storm will dump its rain in the next half hour no longer has to require a supercomputer, according to researchers at the University of São Paulo (USP). Their new tool, TITAN-LSTM, combines weather radar data with artificial intelligence to produce very short-term forecasts, updating a storm’s path, area and rainfall intensity every five minutes, Estadão reported in an Agência Fapesp article published on October 5.

The work was led by Andrea Salome Viteri López, a postdoctoral researcher at USP’s Institute of Astronomy, Geophysics and Atmospheric Sciences (IAG), with support from Fapesp, and was published in the Journal of Geophysical Research: Machine Learning and Computation. The name joins two generations of technology: TITAN, a 1990s algorithm that identifies and tracks individual storm cells on radar, and an LSTM, a neural network designed to remember sequences. TITAN says where a storm is and where it is heading; the network estimates how it will grow, weaken or change shape.

The team trained the system on data collected between 2016 and 2019 by the radar at the Ponte Nova dam in Biritiba Mirim, in greater São Paulo, which scans every five minutes over a 120-kilometre radius. From more than 32,000 storms, they kept 439 that lasted over 20 minutes without merging with other cells, using 307 for training and 132 for testing. In the first minutes of a forecast the hit index rose from 0.2 to about 0.7 compared with the traditional method, and false alarms fell by almost 70 percent. The rainfall intensity error was around 3 millimetres per hour five minutes ahead and roughly 3 to 4 millimetres per hour at 30 minutes.

What it means in practice

The practical appeal is cost. Training the full configuration took about 20 minutes on an ordinary computer using Python, and López says the software can be installed on any laptop and adapted to other radars. That is a very different proposition from the heavyweight infrastructure the country is planning, such as the LNCC’s billion-real AI supercomputer in Macaíba, and it fits the push for universities to share data for applied machine learning.

For civil defence teams, a reliable 15- or 30-minute warning about where heavy rain will fall can inform decisions such as closing roads or moving people out of risk areas, López said.

There are limits. For now the tool handles only “continuous” storms that can be followed as a single object, not cells that merge or split, and the team says it is already working on those more complex cases. What to watch: tests on other Brazilian radars, and whether the forecasts start reaching the agencies responsible for disaster alerts.