Periodic Reporting for period 1 - AQplus4 (Deep Learning Air Quality Forecasts for Four Days)
Reporting period: 2023-11-01 to 2025-10-31
Air pollution poses significant risks to respiratory health and ecosystem stability, yet current operational forecasting systems, such as those used by the Copernicus Atmospheric Monitoring Service (CAMS), are often costly to operate and struggle to capture local pollution patterns accurately. AQPlus4 set out to address this challenge by developing a forecasting system based on advanced artificial intelligence—specifically, Transformer models originally designed for language data. The goal was to create accurate, reliable, and low-cost forecasts that can be adapted globally.
The project focused on ground-level ozone (O₃) and leveraged over 30 years of observations from the Tropospheric Ozone Assessment Report (TOAR) Phase II database alongside ERA5 meteorological data from the European Centre for Medium Range Weather Forecasts (ECMWF). The prototype was built using a Temporal Fusion Transformer (TFT) architecture capable of learning long-range patterns and incorporating anthropogenic factors (such as land use and population density) that traditional models often overlook.
Building on the ERC Advanced Grant IntelliAQ, the project aimed to validate a scientifically grounded, operation-ready forecasting system capable of 4-day predictions. A key objective was to demonstrate the transferability of the model to new regions, specifically evaluating its performance in South Korea using minimal local data. Furthermore, the project developed the technical foundations for automated operational workflows, creating a blueprint for transitioning research-grade models into scalable, service-oriented applications which are affordable also for less affluent countries.
The project focused on ground-level ozone (O₃) and leveraged over 30 years of observations from the Tropospheric Ozone Assessment Report (TOAR) Phase II database alongside ERA5 meteorological data from the European Centre for Medium Range Weather Forecasts (ECMWF). The prototype was built using a Temporal Fusion Transformer (TFT) architecture capable of learning long-range patterns and incorporating anthropogenic factors (such as land use and population density) that traditional models often overlook.
Building on the ERC Advanced Grant IntelliAQ, the project aimed to validate a scientifically grounded, operation-ready forecasting system capable of 4-day predictions. A key objective was to demonstrate the transferability of the model to new regions, specifically evaluating its performance in South Korea using minimal local data. Furthermore, the project developed the technical foundations for automated operational workflows, creating a blueprint for transitioning research-grade models into scalable, service-oriented applications which are affordable also for less affluent countries.
Major scientific progress was achieved despite the unforeseen necessity to find, hire and train new personnel. Work focused on developing a Temporal Fusion Transformer (TFT) based ozone forecasting system. The project developed a prototype that delivered more accurate 4-day forecasts than the Copernicus CAMS ensemble, particularly in suburban and urban areas. Key achievements included the provision of probabilistic forecasts and the integration of meteorological, historical, and static anthropogenic data. Ablation studies confirmed that longer historical contexts (336 hours) and meteorological drivers are critical for skill beyond 48 hours, while static metadata allows for site-specific calibration of the model.
Regarding geographic transferability, a model trained on Germany was successfully adapted to South Korea by retraining only metadata components. Performance declined by only 5-10%, demonstrating that the learned atmospheric relationships generalize across different emissions and climate regimes. This capability addresses the needs of regions lacking long-term observational records. Finally, the team developed foundations for operational workflows by defining data ingestion and inference automation protocols, successfully engaging with international partners to discuss deployment strategies.
Regarding geographic transferability, a model trained on Germany was successfully adapted to South Korea by retraining only metadata components. Performance declined by only 5-10%, demonstrating that the learned atmospheric relationships generalize across different emissions and climate regimes. This capability addresses the needs of regions lacking long-term observational records. Finally, the team developed foundations for operational workflows by defining data ingestion and inference automation protocols, successfully engaging with international partners to discuss deployment strategies.
AQPlus4 produced several results pushing the frontier of air-quality forecasting:
1. First benchmark of a general-purpose transformer model against an operational European forecast system
For the first time, a general-purpose transformer architecture (TFT) was employed for air quality forecasting and was able to outperform the CAMS regional ensemble with minimal domain-specific engineering.
This establishes a new paradigm where data-driven models can supplement or even replace conventional and computationally expensive chemistry transport models (CTMs) for regulatory forecasting tasks.
2. First demonstration of cross-regional transferability for medium range ozone forecasting
The successful transfer of a European-trained model to South Korea shows that learned atmospheric relationships can generalize beyond regional emission and climate regimes.
Hence, the use of deep learning may drastically reduce the cost of forecasting systems for data-sparse regions.
3. First uncertainty-aware multi-day probabilistic ozone forecasts using deep learning
The project introduced distributed (quantile) forecasts, enabling agencies to assess risk and uncertainty.
4. Identification of key requirements for operational deployment
The project clarified several needs for successful long-term uptake:
• scalable data infrastructures (TOAR-II, ERA5) with performant interfaces and high availability.
• future meteorological forecasts must be integrated into the operational workflow.
• metadata quality and availability strongly affect model portability.
• both scientific and engineering components must be combined into a robust integrated system.
Overall, AQPlus4 substantially advanced the technical readiness of deep-learning forecasting systems showing their potential to transform AQ management globally.
1. First benchmark of a general-purpose transformer model against an operational European forecast system
For the first time, a general-purpose transformer architecture (TFT) was employed for air quality forecasting and was able to outperform the CAMS regional ensemble with minimal domain-specific engineering.
This establishes a new paradigm where data-driven models can supplement or even replace conventional and computationally expensive chemistry transport models (CTMs) for regulatory forecasting tasks.
2. First demonstration of cross-regional transferability for medium range ozone forecasting
The successful transfer of a European-trained model to South Korea shows that learned atmospheric relationships can generalize beyond regional emission and climate regimes.
Hence, the use of deep learning may drastically reduce the cost of forecasting systems for data-sparse regions.
3. First uncertainty-aware multi-day probabilistic ozone forecasts using deep learning
The project introduced distributed (quantile) forecasts, enabling agencies to assess risk and uncertainty.
4. Identification of key requirements for operational deployment
The project clarified several needs for successful long-term uptake:
• scalable data infrastructures (TOAR-II, ERA5) with performant interfaces and high availability.
• future meteorological forecasts must be integrated into the operational workflow.
• metadata quality and availability strongly affect model portability.
• both scientific and engineering components must be combined into a robust integrated system.
Overall, AQPlus4 substantially advanced the technical readiness of deep-learning forecasting systems showing their potential to transform AQ management globally.