Google Research and Google DeepMind have developed MetNet-3, an AI-powered weather forecasting model that provides accurate forecasts up to 24 hours in advance.
The model can predict various aspects such as rain, temperature, wind speed and direction, and dew point.
According to the research team, MetNet-3 is more capable than traditional numerical weather prediction (NWP) models such as High-Resolution Rapid Refresh (HRRR) and Ensemble Forecast Suite (ENS) when it comes to forecasts for multiple regions and up to 24 hours in advance.
Real-time AI weather forecasting
A unique feature of the MetNet models is that they use direct observations of the atmosphere for training and evaluation, rather than relying on data from other weather prediction models (NWPs). MetNet-3 uses measurements from weather stations as input and targets, and attempts to produce forecasts for all locations.
The main innovation of the model is a technique called densification, which combines data assimilation (using real data) and simulation in a single step within the neural network. MetNet-3 predicts a probability distribution for each variable and location, providing more information than just the mean.
According to the research team, densification combined with lead-time conditioning and high-resolution direct observations enables a detailed 24-hour forecast with a temporal resolution of two minutes.
MetNet-3’s performance in predicting precipitation over the entire 24-hour period is better than ENS, according to the Google team, as measured by the Continuous Ranked Probability Score (CRPS), a metric of forecast accuracy.
Better weather features in Google products
According to Google, MetNet-3’s real-time system is capable of providing precipitation forecasts every few minutes for the entire U.S. and 27 countries in Europe for up to 12 hours.
The system uses up-to-date data from multiple sources, creates forecasts, checks their accuracy, learns from model results, and makes the results available to millions of Google users.
For example, users searching for weather information for a specific location on their mobile devices will receive highly localized rain forecast data, including detailed timeline graphics with minute-by-minute breakdowns.
Currently available in the US and parts of Europe, MetNet-3 focuses on 12-hour rainfall forecasts.
According to Google, the model is used in various Google products and technologies that process weather information and has the potential to “improve the safety and efficiency of many activities, such as transportation, agriculture, and energy production.”