Fine-scale air quality forecasting in the Grand Est region using machine learning models

ATMO

ATMO Grand Est is the reference expert on air quality in the Grand Est region. As a public interest organization, it brings together all regional stakeholders involved in air monitoring, public communication, and the implementation of actions to improve air quality.

Accueil 

Technical/scientific Challenge:

ATMO Grand Est aimed to develop a forecasting tool capable of simulating pollutant dispersion on an hourly basis across a test territory. The challenge was to implement AI algorithms that could provide high-resolution predictions efficiently. To achieve this, the organization required high-performance computing resources specifically suited to machine learning, enabling rapid scenario testing and forecasting while maintaining accurate results.

Solution:

To meet these needs, the deterministic calculation engine, based on solving physical equations, was replaced with a neural network approach. This shift significantly increased processing speed while maintaining comparable performance. The solution involved generating a representative dataset covering a full year, initially produced from a deterministic model simulating 50 days of hourly data. This dataset was then used for training, optimizing, and evaluating the neural network.

The forecasting tool was installed on a GPU-equipped computing center and deployed on a test domain in March 2022. The first version utilized a Convolutional Neural Network (CNN), allowing fine-scale consideration of road emissions. A second version, based on a U-NET architecture, incorporated both road and surface emissions, used an updated training dataset (Sirane), and expanded the learning domain. These advancements enabled significant optimization of forecasting time and paved the way for deployment across the entire Grand Est region (20 domains) at the beginning of 2023.

Business impact:

The implementation of this AI-based forecasting tool allowed ATMO Grand Est to produce rapid and accurate air quality predictions at a resolution of 25 meters. It provided the capability to test the impact of actions on air quality quickly, supporting more informed decision-making for environmental management. The approach also reduced computational costs and time, ensuring the system could be operated efficiently at scale for the entire region.

Benefits:

  • Full forecasting workflow implemented, including a backup chain
  • Computational time for one zone: 7 hours on 1 CPU (reduced to 30 minutes on 20 CPUs) | Sirane
  • Computational time with 1 GPU: 5–10 minutes | SiraNET
  • Significant acceleration of scenario testing while maintaining high prediction accuracy