Google AI has introduced MetNet-2 – an improved version of the MetNet weather prediction model. MetNet-2 allows you to forecast the weather 12 hours ahead with a spatial resolution of 1 km and a time resolution of 2 minutes.
The application of deep learning models to weather forecasts is in demand not only for use by ordinary users but also in such areas as food production, transport, and energy. Weather forecasts are usually based on physical models and calculated using the largest supercomputers. This approach is limited by high computational requirements and the sensitivity of forecasts to the physical approximations used.
Deep learning offers a new approach to calculating weather forecasts. Instead of explicitly taking into account physical laws, deep learning models learn to predict weather conditions directly based on observed data and work faster than traditional methods, as well as have the potential to increase the accuracy of forecasts.
The first version of MetNet based on radar data and satellite images made forecasts for a maximum of 8 hours ahead. The new model uses 4 times more input data, using data on an area of 64×64 km in the vicinity of the area for which the forecast is generated. The model also takes into account the pre-calculated temperature, humidity and wind direction at the initial moment of time, similar to how it is done in physical models.
MetNet-2 forecasts are superpositions of all possible weather conditions with weights corresponding to the probability of their realization. Because of its probabilistic nature, MetNet-2 can be compared to physical ensemble models. MetNet-2, however, builds a forecast in about 1 second, and ensemble models in 1 hour.