CE04 - Innovations scientifiques et technologiques pour accompagner la transition écologique

Sensor Augmented weather Prediction at HIgh-Resolution – SAPHIR

Sensor Augmented weather Prediction at High Resolution

Short-term, highly localized prediction of atmospheric variables based on deep learning that leverages physical model outputs and data from a network of stations.

Localized forecasting of atmospheric variables based on a hybrid Simulation/Data approach

Providing reliable forecasts of severe weather events is a major issue in many areas such as civil safety or renewable energy production. SAPHIR proposes to combine high-resolution (sub-km) atmospheric dynamics models and a set of direct measurements from weather stations, atmospheric monitoring programs or a dedicated sensor network within a «deep learning” architecture specifically optimized for improving forecasting accuracy. We plan to use this approach to improve the forecast (at horizons ranging from few hours to few days) of intense weather events including rainfalls and electrical activity with an application to river flooding forecasting. SAPHIR is also aiming for an application in the field of renewable energies by improving the forecast of cloudiness and wind strength, which are determining factors for the production of solar or wind power plants.

Our approach is based on the use of AI methods and more generally machine learning methods to optimize the integration of atmospheric model outputs and measurements from a number of sensors in order to best predict a wide range of atmospheric variables at a given site (cumulative precipitation, electrical activity, wind speed, solar radiation, etc.).

The learning architectures we have started to experiment with are based on «deep« neural networks (yet simple) that process temporal information through recurrent networks and spatial information through convolutional networks.

The first work within the framework of the project made it possible to establish the following two highlights:
a- The local and short-term forecasts of statistical models are significantly improved by taking into account information of a spatio-temporal nature such as data measured in neighboring sites.
b- The short and medium term forecasts of precipitation accumulation obtained from a numerical model of the evolution of the atmosphere (such as the Arome model of Météo France) are not very reliable when viewed at a localized site. In particular, these forecasts may turn out to be less good than those given by very simple statistical models. Hybrid approaches that use both data from a numerical model and station measurements are the most effective.

The perspectives arising from our initial results are to demonstrate that the approach envisaged by SAPHIR significantly improves the quality of forecasts compared to approaches solely based on atmospheric dynamic simulation models. Initially, we aim to confirm the relevance of our models in predicting rainfall accumulations at a specific site within time horizons ranging from a few hours to a day. Subsequently, the performance of these models will be evaluated in severe weather events (storms, intense thunderstorms, ...).

To date, two articles in international journals and a DMP have been published.

Providing reliable forecasts of severe weather events is a major issue in many areas such as civil safety or renewable energy production. SAPHIR proposes to combine high-resolution (sub-km) atmospheric dynamics models and a set of direct measurements from weather stations, atmospheric monitoring programs or a dedicated sensor network within a "deep learning” architecture specifically optimized for improving forecasting accuracy. We plan to use this approach to improve the forecast (at horizons ranging from few tens of minutes to few days) of intense weather events including rainfalls and electrical activity with an application to river flooding forecasting. SAPHIR is also aiming for an application in the field of renewable energies by improving the forecast of solar irradiance and wind strength, which are determining factors for the production of solar or wind power plants.
The 4-year project will explore the potentialities of these approaches to improve predictions of intense weather events in the French Mediterranean area and in particular in Corsica region. The Mediterranean basin has quite a unique character with specific physiographic conditions with Corsica featuring a unique observational set-up of an island with urbanized littorals, high mountains and numerous rivers.
SAPHIR is organized in four tasks. The first one is devoted to the problem of collecting and storing weather and environmental data on which our approach relies. Our ambition is to make these data open, accessible and validated for each application task. The second task aims at designing the numerical and software framework in order to perform the prediction of various weather variables, with an extensive use of open source deep learning frameworks and high-performance daily computation of a limited-area numerical weather prediction model. The third task is devoted to applying this methodology to the forecasting, the occurrence, location, timing and intensity of thunderstorms. This prediction will be then exploited in a numerical model in order to predict river flooding in a catchment area. The last task is an application to renewable energy production that consists anticipating and mitigating the production of solar power plants or wind turbine farms for the management of renewable energy facilities. Special attention will be paid to the prediction of extreme wind speed or solar irradiance events. A separated coordination task is also defined to monitor schedule, collaboration, organize communication and ensure production of deliverables and reports.
Overall SAPHIR relies on a solid experience of the consortium members in the fields of high- resolution simulation (SPE, LA, INRIA), forecasting renewable energy resource (SPE) and Mediterranean storm studies (LA). We will also benefit from various experimental material and equipments available at SPE laboratory and from a large observation network and platforms already deployed in Corsica.

Project coordinator

Monsieur Jean-François MUZY (UMR SCIENCES POUR L'ENVIRONNEMENT)

The author of this summary is the project coordinator, who is responsible for the content of this summary. The ANR declines any responsibility as for its contents.

Partner

S P E UMR SCIENCES POUR L'ENVIRONNEMENT
LAERO Laboratoire d'aérologie
INRIA Centre de Recherche Inria de Paris

Help of the ANR 295,972 euros
Beginning and duration of the scientific project: January 2022 - 48 Months

Useful links

Explorez notre base de projets financés

 

 

ANR makes available its datasets on funded projects, click here to find more.

Sign up for the latest news:
Subscribe to our newsletter