As humans, we have learned that there are limitations in what we can predict the atmosphere will do. It’s a chaotic system with many tiny details that influence how the atmosphere evolves.
Thus, humans developed computer forecast models – complex computer programs that run on the world’s fastest supercomputers and solve the intricate fundamental equations of motion, atmospheric dynamics, fluid mechanics, etc. that can convert current observations of the atmosphere – which also have inherent errors – into future predictions of how the atmosphere might evolve through time. These models prove to be relatively accurate with short-term weather forecasts and continue to improve, becoming more accurate with medium- and long-range forecasts.
Recently, there have been attempts to use Deep learning technology to produce weather forecasts. Deep learning is the primary technology behind driverless cars, face recognition, and voice control; it teaches computers to learn by example. This comes naturally to humans of course, as we learn from a young age how to classify tasks, recognize various patterns, and so on. Computers don’t have this natural capability, but if they could be taught, could they do things like recognize weather patterns based on past weather data easier and more quickly than humans would be able to? Would this be any more or less accurate than numerical computer models that use defined physical equations?
Jonathan Weyn was the lead author on a study that found that the Deep learning technique they used was not as accurate as numerical weather prediction models within a few weeks in advance but did relatively well 4-6 weeks in advance. The primary advantage of this model was that it was very quick, taking only 3 minutes to complete a full 320-member ensemble forecast. The full article is found here.
With our deep experience with such an expansive data set as ForecastWatch, we can help evaluate the accuracy of machine learning used in weather forecasting systems. Contact us for more information.