IT Home reported on May 24 that the Microsoft Start team published a blog today through the official blog of Bing, announcing that the AI weather prediction model has been significantly improved, which can more accurately predict the weather conditions in the next 30 days, Now predict when clouds and precipitation may occur.
In the blog, the Microsoft Start team said that it had started to improve the global "precipitation forecast" system at the end of 2021.
This model will not only call the data collected by the radar device, but also combine with the satellite data. Through AI algorithm analysis, it can more accurately predict the precipitation.
However, the Microsoft Start team also admitted that the prediction model was insufficient. The daily receiving time ratio of satellite weather data was 85% to 95%, depending on the data source and acceptable delay time.
Microsoft said:
There is evidence that each task needs a separate decoder, and each prediction channel needs a separate discriminator. Therefore, the model built by Weather from Microsoft Start is four times larger than the previous model that only predicted simulated radar reflectivity. Finally, the new model predicts the reflectivity of satellite and simulated radar jointly, so that its prediction can fill the gap of data availability.
According to Microsoft, in AI training, the weight of radar channel model is six times that of satellite model, because the precipitation information provided by radar is considered more important than satellite tasks. The final result is that the new precipitation and cloud model "has a significant improvement in F1 scores" compared with the basic model only using radar prediction.
Reference address attached to IT home