Microsoft develops an AI model that can provide highly accurate 30-day weather forecasts

In March 2023, weather forecast monitoring company WeatherWatch awarded the Microsoft Start weather software development team the title of "World's most accurate global hydrometeorological forecast service provider". It seems the Microsoft Start team isn't resting on its laurels. This week, Microsoft announced that the company's engineers are developing an AI model that can provide highly accurate weather forecasts 30 days in advance.

In a post on the Bing blog, Microsoft said the company's Start weather software development team published a paper on the new mid-range forecast model on the arXiv site at Cornell University. The article shows how this new model compares with the current model used by the European Center for Medium-Range Weather Forecasts (ECMWF). In short, Microsoft's new system combines the use of up to five AI models and three different deep learning architectures to provide weather forecasts over a period of one month with a high level of accuracy.

Microsoft develops an AI model that can provide highly accurate 30-day weather forecasts Picture 1

Microsoft says these AI models have been trained on massive weather data collected over decades to compile their own prediction methods:

Our AI models work in a similar way to the NWP model: based on the current state of the atmosphere on a 3-D globe (latitude, longitude and altitude), predict the state of the atmosphere atmosphere at some future time, such as one hour later. They then fed this prediction back into the model to predict two hours later, etc

Microsoft's AI models can run predictions faster when using GPUs, and so they can be deployed more frequently. In theory, that should lead to better forecasts. Advanced application makes it possible for AI to quickly sift through troves of weather and climate data, generate thousands of potential outcomes, and finally calculate the probability of certain weather outcomes occurring in a location. specific point.

So far, Microsoft's internal testing has shown that its methods of measuring temperature errors are up to 17% more accurate than the method used by ECMWF for one-week weather forecasts, and 4% more accurate for forecasts. Weather reports continuously for four weeks. The company will add this 30-day weather forecast model to its Microsoft Start model.

Lesley Montoya

Update 07 May 2024