Physics-Informed Neural Networks for Radar-Based Precipitation Nowcasting
Bidragets beskrivning
Reliable short-range forecasts (nowcasts) of rainfall are critical for the society. For instance, rapidly updated high-resolution nowcasts are needed for predicting flash floods in densely populated urban areas. The project aims at using machine learning methods for this purpose. The problem is approached by using interpretable neural networks that, in addition to forecasts, provide information about the underlying physical phenomena and the forecast uncertainties. The neural network utilizes prior physical information via differential equations describing the state of the atmosphere. In addition, it incorporates a statistical cell life cycle model and microphysical information obtained from radar-measured polarimetric parameters. Compared to the existing models, this leads to higher accuracy and computational efficiency. The input data is obtained from the weather radar networks deployed in Finland and the United States, which is complemented by numerical weather prediction models.
Visa merStartår
2021
Slutår
2025
Beviljade finansiering
Övriga uppgifter
Finansieringsbeslutets nummer
341964
Vetenskapsområden
Geovetenskaper
Forskningsområden
Meteorologia ja ilmakehätieteet, ilmastotutkimus
Identifierade teman
computer science, information science, algorithms