What question did this study set out to answer?

The main aim is to enhance the prediction of ozone concentrations by integrating VOC measurements and using dimensionality reduction techniques.

February 17, 2026Open Access

Ozone concentration prediction using machine learning models integrated with 115 volatile organic compound measurements and dimensionality reduction techniques

Key Points

The main aim is to enhance the prediction of ozone concentrations by integrating VOC measurements and using dimensionality reduction techniques.
Utilized VOC monitoring data from 115 species.
Evaluated four machine learning models: LSTM, XGBoost, RF, and LightGBM.
Implemented a novel dimensionality reduction module called SARFE.
Conducted feature importance analysis to identify key predictors.
SARFE-LSTM model achieved R^2 values between 0.78 and 0.92.
Mean absolute errors ranged from 11.02 μg/m^3 to 18.36 μg/m^3.
Root mean square errors varied from 14.73 μg/m^3 to 23.69 μg/m^3.
Improvements in predictive accuracy were observed, with R^2, MAE, and RMSE enhancing by significant margins.

Abstract

High-quality prediction of ozone concentrations is essential for formulating effective environmental policies. The integration of monitoring data of volatile organic compounds (VOCs), key precursors in ozone formation, through dimensionality reduction techniques can substantially improve predictive accuracy. This study utilized monitoring data from 115 VOC species and evaluated four machine learning models-Long Short-Term Memory (LSTM), eXtreme Gradient Boosting (XGBoost), Random Forest (RF), and Light Gradient Boosting Machine (LightGBM)-to assess the influence of VOC monitoring data and dimensionality reduction on ozone prediction. Early attempts to directly utilize the full VOC monitoring dataset yield negligible improvements in model performance. A novel module, termed SARFE, integrating the Statewide Air Pollution Research Center (SAPRC) mechanisms and the Recursive Feature Elimination (RFE) algorithm was proposed for dimensionality reduction, significantly enhancing all models' performance. Among these, the SARFE-LSTM model demonstrated superior performance, achieving coefficients of determination (R2) ranging from 0.78 to 0.92, mean absolute errors (MAE) ranging from 11.02 μg/m3 to 18.36 μg/m3, and root mean square errors (RMSE) ranging from 14.73 μg/m3 to 23.69 μg/m3 for 2-h, 4-h, and 6-h predictions. Compared to traditional LSTM, SARFE-LSTM exhibited substantial improvements in predicting high ozone concentrations, with R2, MAE, and RMSE improving by 36.54%-200.00%, 5.66%-18.62%, and 9.42%-22.45%, respectively. Feature importance analysis identified historical ozone levels (6.78%) and isoprene (5.98%) as key predictors. This study highlights the importance of leveraging VOC monitoring data through sophisticated dimensionality reduction techniques to improve ozone prediction accuracy in machine learning models, especially for high ozone concentrations.

Read Full Paperexternally

Bookmark

View Full Paper

Cite This Study

Han et al. (Sun,) studied this question.

synapsesocial.com/papers/699405254e9c9e835dfd5f03 https://doi.org/https://doi.org/10.1016/j.ecoenv.2026.119868

Bookmark

View Full Paper