Nowcasting heavy rain, using machine learning to improve short-term rainfall predictions
What is nowcasting?
Climate change is making heavy rainfall more intense and more frequent. These rainfall events may cause flooding in cities, polders or other regional water systems in The Netherlands. Numerical weather predictions do not sufficiently capture the severe, sudden and local summer precipitation, often occurring during thunderstorms. Nowcasting provides short-term forecasts (0 to 6 hours ahead) at higher spatial and temporal resolution, capturing these events better. To limit computational time, nowcasting is done using simplified physics or machine learning.
Physics-informed generative AI
Recent breakthroughs in deep learning are transforming the field. One of the most promising models is NowcastNet , which uses introduces a physics-informed generative model that combines data-driven learning with physically plausible motion fields. It outperforms existing methods by producing both realistic and temporally consistent forecasts, and supports ensemble generation, making it suitable for probabilistic forecasting and uncertainty estimation. HKV and Deltares are already running an operational version of NowcastNet. This provides a strong foundation for further research and experimentation.
Your contribution
During your MSc thesis, you will explore how to improve the nowcast further. You’ll apply tools from various disciplines in applied mathematics to resolve these questions. Potential directions for innovation could be to incorporate more physical characteristics or to generate probabilistic nowcasting.
Example research question
How can physical variables and ensemble-based uncertainty quantification improve the performance and usefulness of deep learning precipitation nowcasts for high-intensity rainfall in the Netherlands?
Collaboration and supervision
The project will be supervised by Yuliya Shapovalova from Radboud University. You’ll be also supervised by experts from HKV Lijn in Water and Deltares, and become part of a broader national effort to develop operational nowcasting tools for Dutch water authorities. Office space is available in Delft, Amersfoort or Lelystad, and you’ll get to join meetings with Dutch waterboards and other users — helping your ideas have direct practical relevance.
Contact: Yuliya Shapovalova
References:
Zhang, Y., Long, M., Chen, K., Xing, L., Jin, R., Jordan, M. I., & Wang, J. (2023). Skilful nowcasting of extreme precipitation with NowcastNet. Nature, 619(7970), 526-532