On the 11th of January, I attended a webinar conference organized by IEA Wind Task 51 on Machine Learning Weather Prediction.
The talk has been recorded and is available on YouTube. The three-hour presentation was given by four researchers : - Mariana Clare from ECMWF on "The rise of data driven weather forecasting"
- Gregory Hakim from the University of Washington on "Dynamical Test of a Deep-Learning Weather Prediction Model"
- Joel Oskarsson from Linköping University on "Neural Weather Prediction for Limited Area Modelling"
- Florian Achermann from ETH and JUA.ai on "Ultra High Resolution Wind
Forecasts"
The
first presentation from Mariana Clare gives an overview of the current
state of the work done by ECMWF on data driven weather forecasting.
Since
2022, ten AI models have been developed by different companies and/or
universities. Since then, the European Centre for Medium-Range Weather
Forecasts have run deep analysis on a few chosen AI models:
The
test is split in two categories. First, the skill of each model is
compared with the classical IFS HRES model from ECMWF. The results for
different variables are shown below.
|
A
scoreboard represent the skill of a given forecast compared to the IFS
HRES model (state of the art for physics-based models) (source : Blog)
|
FourCastNet
is not shown here, but this image highlights how well PanguWeather, but
mostly, GraphCast does compare to the ECMWF model.
Skill score
are one thing when assessing the performance of weather models, but
extreme events are even more important. The next part of the test is
then to analyse certain past events and see how well AI model are able
to depict them.
Dr Clare, proceeds on looking at two particular
events : Storm Eunice, which hit Great Britain in 2022 and the Cyclone
Freddy which occurred in the Indian Ocean. The different models (IFS
HRES, FourCastNet, PanguWeather, and GraphCast) are all able to predict
the storm 60 hours in the future. However, AI models (PanguWeather and
FourCastNet) do not show the same results for the prediction of the
Cyclone 48 hours in the future compared to the HRES model.
The AI
models assessed are all trained using 40 years of reanalysis data from
ERA5 using multiple variables (For GraphCast the trained data set is
approximately 35GB). Then, the models are fine-tuned using a more
accurate analysis for specific time periods. As the reanalysis dataset,
the models have a spatial resolution of 0.25°.
The
main advantages of AI models are the time and energy consumption of one
forecast. After being trained, AI models are 500,000 less time and
energy consuming than a classic HRES forecast. ECMWF have developed
python modules for each of the three AI models which can be installed on
a local machine :
- pip install ai-models-panguweather
- pip install ai-models-fourcastnetv2
- pip install ai-models-graphcast
While
analysing these different models, ECMWF also wanted to develop their
own using a similar approach to the DeepMind model. The first version
seems to show good results, but their main goal is to increase the
spatial resolution.
She finishes the talk by mentioning the challenges in the years to come for AI models :
- Train the models on observation as well the reanalysis.
- Move from deterministic to probabilistic models to integrate measures of uncertainty.
- Build trust in AI models for researcher.
The next talk by Gregory Hakim gives the detailed process of testing a deep learning weather model.
Joel Oskarsson then, gives an explanation of how AI models can run on a limited area.
Florian Achermann finishes by explaining the development of a high resolution wind forecasting model.