Accurate weather forecasting plays a vital role in our daily lives, influencing everything from planning outings to energy production. With the increasing occurrence of extreme weather events like floods, droughts, and heatwaves, the need for precise forecasts has never been more crucial. In particular, the first 24 hours of weather predictions are immensely valuable as they are both highly predictable and actionable, aiding individuals in making informed decisions promptly to ensure safety.

MetNet-3
MetNet-3

Today, we introduce MetNet-3, a revolutionary weather model developed jointly by Google Research and Google DeepMind. Building upon the successes of our earlier models, MetNet and MetNet-2, MetNet-3 is designed to offer high-resolution predictions for up to 24 hours into the future, encompassing a broader range of core variables including precipitation, surface temperature, wind speed and direction, and dew point. MetNet-3 delivers a temporally smooth and highly detailed forecast, offering lead time intervals as short as 2 minutes and spatial resolutions ranging from 1 to 4 kilometers. Our model excels when compared to traditional methods, even outperforming the best physics-based numerical weather prediction models such as High-Resolution Rapid Refresh (HRRR) and ensemble forecast suite (ENS) for multiple regions in the 24-hour prediction window.

What’s more, we’ve seamlessly integrated the capabilities of MetNet-3 into various Google products and technologies where weather information is pertinent. Presently, this service is available in the contiguous United States and select parts of Europe, with a focus on delivering 12-hour precipitation forecasts. MetNet-3 is playing a pivotal role in providing precise and reliable weather data to users across multiple countries and languages.

 

Densification of Sparse Observations

While many recent machine learning weather models rely on atmospheric data generated by conventional methods, the MetNet series has stood out by using direct observations of the atmosphere for training and evaluation. These direct observations offer superior fidelity and resolution. However, direct observations come from various sensors at different altitudes, including surface-level weather stations and orbiting satellites, which can vary in terms of data density. For example, radar-derived precipitation estimates from sources like NOAA’s Multi-Radar/Multi-Sensor System (MRMS) offer relatively dense images, while ground-based weather stations providing temperature and wind data are sparse point measurements.

In addition to the data sources employed in previous MetNet models, MetNet-3 now integrates point measurements from weather stations both as inputs and targets, aiming to make forecasts available for all locations. The key innovation in MetNet-3 is a process called densification, which consolidates the traditional two-step approach of data assimilation and simulation found in physics-based models into a single neural network pass. This densification process applies to individual data streams, and the resulting forecasts benefit from all input streams in MetNet-3, including topographical, satellite, radar, and NWP analysis features. Notably, MetNet-3 doesn’t rely on NWP forecasts as default inputs.

High Resolution in Space and Time

One of the major advantages of using direct observations is their high spatial and temporal resolution. For instance, weather stations and ground radar stations provide measurements every few minutes at specific points with 1-km resolution, in stark contrast to the 6-hour intervals and 9-km resolution of the state-of-the-art model ENS. MetNet-3 preserves the concept of lead time conditioning, where the forecast’s lead time in minutes is directly input to the neural network. This approach enables MetNet-3 to efficiently model high-frequency observations, even as brief as 2 minutes. The combination of densification, lead time conditioning, and high-resolution direct observations results in a fully dense 24-hour forecast with a 2-minute temporal resolution, based on learning from just 1,000 points from the One Minute Observation (OMO) network of weather stations scattered across the United States.

MetNet-3 predicts a marginal multinomial probability distribution for each output variable at each location, offering rich information beyond just the mean value. This allows for meaningful comparisons between MetNet-3’s probabilistic outputs and those of advanced ensemble NWP models, such as the ensemble forecast ENS from the European Centre for Medium-Range Weather Forecasts and the High Resolution Ensemble Forecast (HREF) from the National Oceanic and Atmospheric Administration of the US. The probabilistic nature of both models’ outputs allows for metrics like the Continuous Ranked Probability Score (CRPS) to be calculated. Notably, MetNet-3’s forecasts are not only of higher resolution but also more accurate when assessed at overlapping lead times.

Performance comparison between MetNet-3 and NWP baseline for wind speed based on CRPS (lower is better).

In contrast to weather station variables, precipitation estimates are denser as they are derived from ground radar. MetNet-3’s approach to modeling precipitation is akin to that of MetNet-1 and 2 but extends high-resolution precipitation forecasts to a spatial granularity of 1 km over the same 24-hour lead time, delivering superior CRPS values compared to ENS.

 

Performance comparison between MetNet-3 and NWP baseline for instantaneous precipitation rate on CRPS (lower is better).

Delivering Real-time ML Forecasts

Training and evaluating a weather forecasting model like MetNet-3 with historical data is just one aspect of delivering ML-powered forecasts to users. Developing a real-time ML system for weather forecasting involves numerous considerations, including processing real-time input data from various sources, running inference, validating outputs in real-time, deriving insights from the model’s output, and serving the results at scale – all in a continuous cycle, refreshed every few minutes.

Our system stands out for its near-continuous inference, allowing the model to provide constant forecasts based on incoming data streams. This mode of inference is necessary due to the diverse characteristics of the incoming data. The model accepts various data sources as input, such as radar, satellite, and numerical weather prediction assimilations, each with distinct refresh frequencies and spatial and temporal resolutions. The system effectively synchronizes these data sources, enabling the model to provide up-to-date precipitation forecasts for the next 12 hours at a high cadence.

With this process, the model can predict discrete probability distributions. Innovative techniques have been developed to transform this rich output space into user-friendly information that enhances user experiences across Google products and technologies.

Weather Features in Google Products

People worldwide rely on Google for timely and accurate weather information. This information serves various purposes, from planning outdoor activities to ensuring safety during severe weather conditions.

MetNet-3’s exceptional accuracy, high temporal and spatial resolution, and probabilistic nature enable the creation of unique hyperlocal weather insights. In the contiguous United States and Europe, MetNet-3 is operational, delivering real-time 12-hour precipitation forecasts, now available across Google products and technologies where weather information is essential, including Google Search. The model’s output is translated into actionable information, providing millions of users with highly localized precipitation forecasts, complete with minute-by-minute breakdowns.

 

MetNet-3 precipitation output in weather on the Google app on Android (left) and mobile web Search (right).

Conclusion

MetNet-3 represents a groundbreaking deep learning model for weather forecasting that surpasses state-of-the-art physics-based models for 24-hour forecasts of core weather variables. Its potential extends to revolutionizing weather forecasting, benefiting a wide range of activities, from transportation and agriculture to energy production. MetNet-3 is fully operational, with its forecasts seamlessly integrated into various Google products, where accurate weather data is indispensable.

 

Similar Posts

Leave a Reply

Your email address will not be published. Required fields are marked *