Google's DeepMind grows AI system to predict rain within 2 hrs

The weather predictions are driven by powerful numerical weather prediction (NWP) systems. By solving physical equations, NWPs provide essential planet-scale predictions several days ahead. However, they struggle to generate high-resolution predictions for short lead times under two hours. Nowcasting fills the performance gap in this crucial time interval.

Nowcasting is essential for sectors like water management, agriculture, aviation, emergency planning, and outdoor events. Advances in weather sensing have made high-resolution radar data–which measures the amount of precipitation at ground level–available at high frequency (e.g., every 5 mins at 1 km resolution). This combination of a crucial area where existing methods struggle and the availability of high-quality data provides the opportunity for machine learning to make its contributions to nowcasting.

Generative Models for Nowcasting

Focus on nowcasting rain: predictions up to 2 hours ahead that capture the amount, timing, and location of rainfall. An approach known as generative modelling can be used to make detailed and plausible predictions of future radar based on past radar. Conceptually, this is a problem of generating radar movies. With such methods, we can both accurately capture large-scale events, while also generating many alternative rain scenarios (known as ensemble predictions), allowing rainfall uncertainty to be explored. The company used radar data from both the UK and the US in our study results.

They were especially interested in the ability of these models to make predictions on medium to heavy-rain events, which are the events that most impact people and the economy, and we show statistically significant improvements in these regimes compared to competing methods. Importantly, we conducted a cognitive task assessment with more than 50 expert meteorologists at the Met Office, the UK’s national meteorological service, who rated our new approach as their first choice in 89% of cases when compared to widely-used nowcasting methods, demonstrating the ability of our approach to provide insight to real world decision-makers.

By using statistical, economic, and cognitive analyses we were able to demonstrate a new and competitive approach for precipitation nowcasting from radar. No method is without limitations, and more work is needed to improve the accuracy of long-term predictions and accuracy on rare and intense events. Future work will require us to develop additional ways of assessing performance, and further specialising these methods for specific real-world applications.

This is an exciting area of research and the paper will serve as a foundation for new work by providing data and verification methods that make it possible to both provide competitive verification and operational utility. This collaboration with the Met Office will promote greater integration of machine learning and environmental science, and better support decision-making in our changing climate.