Air Quality Prediction: A Systematic Review Of Traditional Methods And Emerging Hybrid Frameworks

Abstract

Air quality is a major global challenge with pollutants like PM_2.5, PM₁₀, NO₂, SO₂, CO, and O₃ being a threat (or upset) to human health, the environment, and ultimately quality of life. The short-term forecasting of pollutant concentrations (for example, PM_2.5 and PM₁₀) can be immensely useful for early warning systems, regulatory actions, and public advisories. Fore- cast modeling has drastically evolved in the last twenty years. Although traditional modeling strategies such as linear autoregressive integrated moving average (ARIMA), multiple linear re- gression (MLR), and generalized additive models (GAM) are easy to use and interpret, they are generally limited in producing estimates of nonlinear interactions, structural breaks, and long- range dependence. Researchers have looked extensively to utilize machine learning and deep learning algorithms (i.e. Random Forests, Support Vector Machines (SVM), Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), Convolutional Neural Networks (CNN), and transformers) in recent years in an effort to overcome these limitations. These methods im- prove spatiotemporal accuracy compared to traditional statistical methods but they do not offer significant improvement in computational burden, interpretability, or timely deployment. The emergence of hybrid models, which utilize statistical, machine learning, and deep learning-type models with a variety of data sources, such as meteorological, satellite, and IoT-based sensing or other sensor data, has improved robustness and scalability. The review presents the evolution of air quality forecasting techniques, contrasts advantages and disadvantages, and underlines implications for public health and policy.