The application of science and technology to predict atmospheric conditions for a specific location and time is known as weather forecasting. For centuries people have tried to guess the weather informally and since the 19th century methodically. Currently, traditional physics-based techniques powered by the world’s largest supercomputers are used to predict the weather. But high computational requirements limit such methods, and they are also prone to approximations of the physical laws on which they are based.
Deep learning can offer a new approach to calculating forecasts. Deep learning has been used to solve a wide variety of crucial problems, including preventing cancer and improving accessibility already. Hence, using deep learning models to predict the weather can be helpful for humans on a daily basis. Deep learning models learn to predict weather patterns directly from observable data, rather than applying explicit physical laws, and can compute forecasts faster than physics-based techniques. These methods have the potential to increase the frequency, scope, and accuracy of predicted forecasts.
Deep learning algorithms have shown promise in weather forecasting for nowcasting, as they predict the weather up to 2-6 hours in advance. Previous research has focused on using direct neural network models for weather data, extending neural forecasts from 0 to 8 hours with the MetNet architecture, generating radar data continuations for up to 90 minutes in advance, and interpreting the weather information learned from these neural networks. Deep learning, on the other hand, has the potential to improve longer-term forecasts.
Google AI has the. released Meteorological Neural Network 2 (MetNet-2) for 12-hour precipitation forecast. MetNet-2 clearly outperforms its predecessor MetNet. MetNet-2 outperforms the state-of-the-art HREF ensemble model for weather forecasting up to 12 hours compared to physics-based models.
MetNet-2 and other neural weather models relate measurements of the earth to the probability of weather events such as rain over a city in the afternoon, gusts of wind of 20 knots or a sunny day. By directly integrating the inputs and outputs of a system, end-to-end deep learning can both streamline and improve quality. MetNet-2 was developed with this in mind and seeks to reduce both the complexity and the total number of steps required to produce a forecast.
Radar and satellite photos, which are also used in MetNet, are among the inputs of MetNet-2. MetNet-2 uses the preprocessed initial state used in physical models as a proxy for this additional meteorological information to capture a more comprehensive snapshot of the atmosphere with information such as temperature, humidity and wind direction – crucial for longer forecasts of up to 12 hours.
Capturing sufficient spatial information in the input photos is one of the main problems MetNet-2 must overcome in order to produce 12 hour forecasts. The team included 40 miles of background in each direction for each additional hour of forecast at the entrance. This results in an input context of 20,482 km2 that is four times larger than that of MetNet. MetNet-2 updates MetNet’s attention layers with computationally more efficient convolution layers due to the size of the input context.
The forecasts were rated using established metrics such as the Continuous Ranged Probability Score, which measures the probabilistic error of a model compared to the ground truth observations. Although MetNet-2 does not use physics-based calculations, it is able to outperform HREF in both low and high rainfall up to 12 hours in the future.
Since MetNet-2 does not rely on hand-made physical equations, the question that arises for its effectiveness is: What types of physical weather relationships does it learn from the data during training?
The most striking conclusion is that MetNet-2 mimics the physics described by the Quasi-Geostrophic Theory, which is used to approximate large-scale weather events. On the scale of a normal high or low pressure system (i.e., the synoptic scale), MetNet-2 has been able to detect changes in atmospheric forces that result in favorable precipitation conditions, which is a critical component of the theory.
MetNet-2 is a step in enabling a new modeling paradigm for weather forecasting that does not rely on manually coding the physics of weather events. Many obstacles remain in the path to fully attaining this goal, including the direct inclusion of more raw data on the atmosphere.