Automatic Generation of Text Weather Forecast with Spatio-Temporal Aggregation

For many years, Meteo-France forecasters have been writing text forecasts for each district in France. In these forecasts, the weather is summarized in time and space. These text forecasts have a large audience within the general public and business customers in France.

As France is divided into 100 districts, writing a forecast for each district several times a day takes a lot of human resources.

That is why, for 15 years, Meteo-France has been developing the AGAT system to automate the writing of these forecasts. But this system was not entirely satisfactory because it was unable to correctly describe the complex weather situations that were evolving both in time and space.

The advent of Machine Learning techniques has recently allowed AGAT to be recast. Its production is now almost indistinguishable from human production. The new version of AGAT is now operational.

This new version of AGAT is built upon a classification of meteorological situations by Machine Learning (unsupervised learning). Each new weather situation is associated to one class for which a standard text forecast has been written. This standard text is then post-processed to complete the first guess, especially for some parameters such as precipitation type and geographical information.

Here is an extract of a text forecast produced by the new version of AGAT (in French, English version follows): « La nuit débute sous un ciel généralement étoilé. Seules quelques rares averses perturbent le calme ambiant sur le Lauragais. En seconde partie de nuit, le temps est sec, mais le ciel se couvre progressivement. »

Translation by Google Translate: "The night begins under a generally starry sky. Only a few showers disturb the calm on the Lauragais. In the second part of the night, the weather is dry, but the sky is gradually covering."

AGAT is now able to tackle very complex situations, even during the winter period in districts with both plains and mountains.