ICML 2019 Expo Talk

July 18, 2021

Expo 2019 Schedule »

How to build hyperlocal weather forecast with machine learning techniques on a global scale

Sponsor: Yandex

Alexander Ganshin (Yandex)


Atmosphere is a competitive source of big data. Numerical weather prediction models produce terabytes of data everyday. Weather forecasts quality dramatically increased nowadays by the computation power of supercomputers and new sources of weather observations. But forecast couldn’t be absolutely correct, they have errors. Weather prediction could be considered as machine learning task, where features are atmospheric parameters from different weather forecast models, and targets are meteorological observation on ground stations. Yandex operates proprietary technology Meteum, which offers accurate hyperlocal weather forecasts all over the world. It is based on combination of traditional numerical weather prediction models and progressive machine learning algorithms. We use CatBoost and output from 5 different weather models to produce accurate global forecast and real-time weather conditions with 2x2 km spatial resolution. Also we introduce nowcasting technology based on neural networks, geostationary satellite imagery and meteorological radars measurements to provide precipitation maps with temporal resolution up to 10 minutes on a global scale. Scalability, accuracy, high spatial and temporal resolution of the models allow to build global weather forecast including an alerting service with push notifications for the Yandex ecosystem products