6-month internship on the contribution of artificial intelligence to precipitation forecasting in Corsica: first results
Ioan Raguet, an engineering student from the National School of Meteorology in Toulouse, did an end-of-training internship at the Aerology Laboratory of the University of Toulouse 3 – Paul Sabatier (supervised by F. Pantillon and D. Lambert) and at the Laboratory Environmental Sciences from the University of Corsica (supervised by R. Baggio, J.-B. Filippi, and J.-F. Muzy). This internship is part of the ANR SAPHIR project (https://saphir.universita.corsica). It is also a continuation of the HyMeX program (https://www.hymex.org) partly dedicated to studying intense meteorological phenomena regularly affecting the entire north-western Mediterranean basin. It confirms the interest in the research activities on meteorology in Corsica carried out in this context for more than ten years, which has enabled the creation of the CORSiCA Platform for Atmospheric Observations (https://corsica.obs-mip.fr).
Figure 1: Work session at the SPE in Corte during Ioan's stay last May. From left to right: Roberta Baggio, Ioan Raguet, Jean-François Muzy
This 6-month internship (from March to August 2022) focuses on the contribution of artificial intelligence to the forecasting of convective precipitating systems in Corsica. It consists of beginning to evaluate a complementary approach to numerical modeling alone and based on artificial intelligence for forecasting convective systems in Corsica.
Initially, the study aims more precisely to test the performance of forecasting the accumulation of precipitation over 3h at the 3h horizon using a deep dense neural network. Meteorological data from the operational network of Météo France surface stations and a set of numerical simulations carried out with the MesoNH model are used as input to the learning system. The results are compared with the results obtained by the meteorological model alone. They are evaluated also according to the type of meteorological situation studied.
Figure 2 shows an example result. This is a comparison of the performance of the deep dense neural network with the single MesoNH meteorological model for the forecast of cumulative precipitation over 3 hours at the 3-hour horizon on each of the stations of the Météo-France network. The two models are evaluated over the period from 2017 to 2019. In this case, the training of the neural network is based solely on data from the surface stations of Météo-France from 2010 to 2017.
Figure 2: Comparison of the forecast performance of 2 models for the forecast of the cumulative precipitation over 3h at the 3h horizon at a given station over the period from 2017 to 2019. In blue: training of the deep dense neural network with the data surface stations of Météo-France. In red: MesoNH alone.
The evaluation based on root mean square error calculations reveals a better performing neural network than MesoNH over the period from 2017 to 2019.
The study then focused on the prediction performance of the neural network for three highly precipitating events in 2019:
- on October 7, 2019, a night storm appeared over the Mediterranean Sea north-east of Corsica
- on September 21, 2019, with large stormy formations in the evening over the Mediterranean Sea west of Corsica
- on August 30, 2019, precipitating event at midday, located on the Corsican relief
The high rainfall activity of these events is shown in Figure 3.
Figure 3. High rainfall activity
Performance evaluations of the neural network on these days show that it is effective for predicting events associated with large-scale atmospheric circulation, such as October 7 or September 21. On the other hand, it encounters difficulties in predicting local events such as that of August 30.
The first results of this new approach are encouraging and will be pursued within the framework of a thesis that will begin next October, financed by the SAPHIR project. The very recent devastating storm of August 18, 2022, confirms the interest in this project.