Multivariable Air-Quality Prediction and Modelling via Hybrid Machine Learning: A Case Study for Craiova, Romania

Area of Science:

• Environmental Science and Engineering
• Atmospheric Science
• Data Science and Machine Learning

Background:

• Poor air quality negatively impacts human health and exacerbates climate change.
• Understanding the localized relationship between climate variations and air pollution is crucial for mitigating health risks.
• Existing models require enhancement for accurate, localized air quality prediction.

Purpose of the Study:

• To develop accurate air quality predictions using a holistic, multivariate modeling approach.
• To investigate the complex associations between meteorological factors (temperature, humidity, pressure) and particulate matter (PM10, PM2.5, PM1) concentrations.
• To assess the correlation between PM concentrations and other pollutants like noise, volatile organic compounds (VOCs), and carbon dioxide (CO2).

Main Methods:

• Utilized five hybrid machine learning models for predicting PM concentrations and the Air Quality Index (AQI).
• Collected high-frequency (1-minute intervals) data over five months from twelve distributed PM sensors in Craiova City, Romania.
• Employed multivariate analysis to identify significant predictor variables and their influence on PM concentrations.

Main Results:

• Machine learning models achieved high prediction accuracy, with R-squared values generally exceeding 0.96 and often approaching 0.99.
• Temperature and air pressure were identified as the most influential meteorological variables for PM concentration prediction, while relative humidity had the least impact.
• PM10 concentrations showed strong correlation with PM2.5 and moderate correlation with PM1; however, PM concentrations were not strongly related to noise, CO2, or VOCs individually.

Conclusions:

• Established novel, highly accurate predictive relationships for PM concentrations and AQI based on identified key meteorological variables.
• Demonstrated the effectiveness of hybrid machine learning models in localized air quality forecasting.
• Indicated that combining non-meteorological factors (noise, CO2, VOCs) with meteorological variables is necessary to improve their predictive power for air quality.