Prediction and explanation for ozone variability using cross-stacked ensemble learning model.

Area of Science:

• Environmental Science
• Atmospheric Chemistry
• Data Science

Background:

• Ozone prediction studies benefit from increased monitoring data on ozone precursors.
• Machine learning offers advanced methods for analyzing complex environmental datasets.

Purpose of the Study:

• To develop and validate a novel cross-stacked ensemble learning model (CSEM) for accurate hourly ozone concentration prediction.
• To evaluate the impact of emission reduction scenarios on ozone levels.
• To propose a comprehensive evaluation index for prediction models.

Main Methods:

• Feature engineering and reconstruction of multi-source data, including meteorological, conventional pollutant, and volatile organic compound (VOCs) data.
• Development of a CSEM integrating Random Forest, Extreme Gradient Boosting Tree, and Long Short-Term Memory (LSTM) base models.
• Cross-stacked integrated training to enhance ensemble model performance.

Main Results:

• The CSEM achieved high prediction accuracy with R² values of 0.94, 0.97, and 0.96 for mild, moderate, and severe pollution, respectively.
• Mean Absolute Errors (MAEs) were 4.48 μg/m³, 5.01 μg/m³, and 8.71 μg/m³ for the respective pollution levels.
• A 20% reduction in VOCs, with no change in NOx, led to ozone reduction in 75.28% of cases and levels below 200 μg/m³ in 15.73% of cases.

Conclusions:

• The developed CSEM demonstrates superior performance in predicting hourly ozone concentrations.
• The study highlights the effectiveness of machine learning for analyzing complex air quality data and informing emission control strategies.
• The proposed evaluation index offers a standardized method for comparing prediction model performance.