Weather forecasts must capture all possibilities, including the worst-case scenario, which is most important for planning.
WeatherNext 2 can predict hundreds of possible weather outcomes from a single starting point. On a single TPU, each prediction takes less than a minute. It would take many hours on a supercomputer using a physically-based model.
Our model is highly skilled and capable of high-resolution predictions on an hourly basis. Overall, WeatherNext 2 outperforms previous state-of-the-art WeatherNext models for 99.9% of variables (temperature, wind, humidity, etc.) and lead times (0-15 days), enabling more useful and accurate forecasts.
This performance improvement is achieved through a new AI modeling approach called Functional Generative Networks (FGN). FGN injects “noise” directly into the model architecture, so the predictions it produces remain physically realistic and interconnected.
This approach is especially useful for predicting what meteorologists call “boundaries” or “seams.” Limits are separate and independent weather factors, such as the exact temperature at a particular location, wind speed at a particular altitude, or humidity. The novelty of our approach is that the model is trained only at these limits. But from that training, you learn how to skillfully predict large, complex, interconnected systems that depend on the “joints” – how all the individual parts fit together. This “joint” forecasting is necessary for the most useful forecasts, such as determining the expected power production across regions affected by high heat or across wind farms.
