Hybrid AI model for PM2.5 forecasting using multiple data processing techniques: A case study in Jordan

Abstract Accurate forecasting of particulate matter with a diameter of less than 2.5 µm (PM 2.5 ) is crucial for public health, especially in regions like Jordan, where air dusty conditions occur frequently. This study focuses on developing various machine learning models to forecast the PM 2.5 concentrations in urban areas of Jordan, specifically in Amman and Zarqa, covering various environments: background, residential, traffic, and industrial. The used dataset was collected from the daily air quality reports from Jordan’s Ministry of Environment over the period of Jan. 2021 until April 2024. In this work, we developed and evaluated different Machine Learning models, including single, combined, and hybrid models. The models incorporated various environmental factors such as temperature, wind speed, humidity, and air pollutants. The developed hybrid model incorporated techniques like Prophet for anomaly detection, MICE for missing data imputation, Random Forest for feature selection, and Singular Spectrum Analysis for trend and seasonality extraction, paired with multiple forecasting techniques. Our evaluation results show that the hybrid model outperformed other models with strong forecasting results for 1-day to 4-day horizon for background PM 2.5 and 1-day to 2-day forecasting for residential, traffic, and industrial areas. The hybrid model’s Mean Absolute Error ranged from 1.6 µg/m3 to 6.6 µg/m3, with Coefficients of Determination (R 2 ) ranging from 0.915 to 0.59, demonstrating the effectiveness and reliability of the model. This research is a foundational step toward building an awareness system against air pollution, addressing a critical gap in environmental management and public health in Jordan.