Over the past two decades, air-quality prediction has moved from a mainly single-model paradigm toward ensemble systems that make explicit use of diversity across models, observations, and data streams. This review connects developments that are often treated separately: chemical transport model (CTM) ensembles, tree-based and hybrid machine learning ensembles, deep learning architectures, physics-informed neural networks, and distributed approaches such as federated learning. Evidence summarized from recent systematic reviews and coordinated modeling initiatives indicates that, within comparable validation settings, ensembles often outperform individual models for PM2.5, PM10, O3, NO2, CO, and SO2 across a broad range of spatial scales and standard error metrics, including RMSE, MAE, and correlation. Operational CTM ensembles, such as the Copernicus Atmosphere Monitoring Service (CAMS) European system with eleven regional models, improve both forecast skill and uncertainty characterization for ozone and particulate matter. In data-driven applications, tree-based ensembles (Random Forest, gradient boosting, XGBoost, LightGBM) and hybrid deep architectures (CNN–LSTM models, attention-based multi-branch networks, graph neural networks) now form a core part of the state of the art for AQI (Air Quality Index) and particulate-matter estimation from structured and multi-source data. Reported performance can be very high on well-structured tabular datasets, with (Formula presented.) values above 0.99 in selected benchmarks and RMSE reductions of 23–45% relative to classical statistical baselines in multi-modal studies; however, these values are not directly interchangeable because pollutant type, prediction horizon, monitoring density, and validation design differ among studies. This review proposes a practical taxonomy of ensemble strategies and uses it to explain why diversity, rather than model count alone, is central to reliable air-quality prediction. Drawing on coordinated European and North American model-evaluation initiatives (AQMEII, HTAP) and on case studies in topographically and meteorologically complex Italian regions (the Po Valley, the Naples metropolitan area, and Campania), we show that effective ensemble design requires a balance among diversity, redundancy, computational feasibility, and interpretability. On the basis of a structured narrative synthesis, the main research gaps concern physics-informed and explainable ensemble frameworks, transferable and adaptive models, standardized benchmarks, severe-pollution-episode forecasting, and scalable distributed architectures. Open questions include how to design compact non-redundant CTM sub-ensembles and how to couple deep learning with chemical-transport physics in next-generation operational systems.
Multi-Model Ensemble Approaches in Air Quality Prediction: A Comprehensive Review from Chemical Transport Models to Hybrid Machine Learning
Chianese E.;Riccio A.
2026-01-01
Abstract
Over the past two decades, air-quality prediction has moved from a mainly single-model paradigm toward ensemble systems that make explicit use of diversity across models, observations, and data streams. This review connects developments that are often treated separately: chemical transport model (CTM) ensembles, tree-based and hybrid machine learning ensembles, deep learning architectures, physics-informed neural networks, and distributed approaches such as federated learning. Evidence summarized from recent systematic reviews and coordinated modeling initiatives indicates that, within comparable validation settings, ensembles often outperform individual models for PM2.5, PM10, O3, NO2, CO, and SO2 across a broad range of spatial scales and standard error metrics, including RMSE, MAE, and correlation. Operational CTM ensembles, such as the Copernicus Atmosphere Monitoring Service (CAMS) European system with eleven regional models, improve both forecast skill and uncertainty characterization for ozone and particulate matter. In data-driven applications, tree-based ensembles (Random Forest, gradient boosting, XGBoost, LightGBM) and hybrid deep architectures (CNN–LSTM models, attention-based multi-branch networks, graph neural networks) now form a core part of the state of the art for AQI (Air Quality Index) and particulate-matter estimation from structured and multi-source data. Reported performance can be very high on well-structured tabular datasets, with (Formula presented.) values above 0.99 in selected benchmarks and RMSE reductions of 23–45% relative to classical statistical baselines in multi-modal studies; however, these values are not directly interchangeable because pollutant type, prediction horizon, monitoring density, and validation design differ among studies. This review proposes a practical taxonomy of ensemble strategies and uses it to explain why diversity, rather than model count alone, is central to reliable air-quality prediction. Drawing on coordinated European and North American model-evaluation initiatives (AQMEII, HTAP) and on case studies in topographically and meteorologically complex Italian regions (the Po Valley, the Naples metropolitan area, and Campania), we show that effective ensemble design requires a balance among diversity, redundancy, computational feasibility, and interpretability. On the basis of a structured narrative synthesis, the main research gaps concern physics-informed and explainable ensemble frameworks, transferable and adaptive models, standardized benchmarks, severe-pollution-episode forecasting, and scalable distributed architectures. Open questions include how to design compact non-redundant CTM sub-ensembles and how to couple deep learning with chemical-transport physics in next-generation operational systems.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
