Comparison of machine learning methods for predicting ground-level ozone pollution in Beijing

Aug 3, 2026

Uploaded by Rowan

High ground-level ozone (O3) concentrations degrade urban air quality and pose risks to human health, making accurate forecasting essential for environmental monitoring and policy decisions. This study built a Lagged Feature Prediction Model by using ozone and NO2 concentrations from the previous three hours as lagged inputs, then evaluated nine machine-learning algorithms, including XGBoost. XGBoost with SHAP selected 11 key features and improved efficiency by 30%. Meteorology-only models performed best with ED-LSTM (R2=0.479), while adding pollutant variables substantially increased accuracy; XGBoost reached R2=0.767 (RMSE=11.35 μg/m3). Incorporating the LFPM further improved results, with XGBoost achieving R2=0.873 (RMSE=8.17 μg/m3).

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages19

Reading Time48 min

AccessFree to download

Introduction

Health and environmental impacts of ground-level ozone

Drivers and photochemical complexity

Existing prediction approaches and limitations

Methods and modeling (inferred)

Lagged Feature Prediction Model (LFPM)

Machine-learning algorithms and feature selection

Evaluation using meteorological and pollutant variables

Results (inferred)

Model comparisons and performance metrics

SHAP-selected key features and efficiency

Conclusion (inferred)

Key findings on lagged feature integration

TYPE Original Research PUBLISHED 01 April 2025 DOI 10.3389/fenvs.2025.1561794 OPEN ACCESS EDITED BY Run Liu, Jinan University, China REVIEWED BY Xiang Weng, University of East Anglia, United Kingdom Luo Zheng, Jinan University, China *CORRESPONDENCE Zhigang Lu, [zhiganglu14@163.com](zhiganglu14@163.com)[ ](zhiganglu14@163.com)Weidong Zhu, [wdzhu@shou.edu.cn](wdzhu@shou.edu.cn) RECEIVED 16 January 2025 ACCEPTED 05 March 2025 PUBLISHED 01 April 2025 CITATION Liu Z, Lu Z, Zhu W, Yuan J, Cao Z, Cao T, Liu S, Xu Y and Zhang X (2025) Comparison of machine learning methods for predicting ground-level ozone pollution in Beijing. Front. Environ. Sci. 13:1561794 . doi: 10.3389/fenvs.2025.1561794 COPYRIGHT © 2025 Liu, Lu, Zhu, Yuan, Cao, Cao, Liu, Xu and Zhang. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY) . The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms. Comparison of machine learning methods for predicting ground-level ozone pollution in Beijing Zitao Liu 1, Zhigang Lu 2*, Weidong Zhu 1,3*, Jiansheng Yuan 1,3, Zhaoxiang Cao 1, Tiantian Cao 1, Shuai Liu 1, Yuelin Xu 1 and Xiaoshan Zhang 1 1College of Oceanography and Ecological Science, Shanghai Ocean University, Shanghai, China, 2School of Resources and Civil Engineering, Gannan University of Science and Technology, Ganzhou, China, 3Shanghai Engineering Research Center of Estuarine and Oceanographic Mapping, Shanghai Ocean University, Shanghai, China High ground - level ozone (O3) concentrations severely undermine urban air quality and threaten human health, creating an urgent need for precise and effective ozone-level predictions to aid environmental monitoring and policymaking.This study incorporated the historical concentrations of ozone and nitrogen dioxide (NO2) from the past 3 hours as lagged features into a Lagged Feature Prediction Model (LFPM), evaluated using nine machine - learning algorithms (including XGBoost) . Initially, XGBoost combined with SHAP identiﬁed 11 key features, boosting computational efﬁciency by 30% without sacriﬁcing prediction accuracy. Then, ozone concentrations were predicted using six meteorological variables. Results showed that LSTM-based methods, especially ED-LSTM, performed best among meteorological-only models (R2 = 0.479) . Yet, predictions based solely on meteorological variables had limited accuracy. Adding ﬁve pollutant variables markedly improved the predictive performance across all machine - learning methods. XGBoost achieved the highest accuracy (R2 = 0 .767, RMSE = 11 .35 μg/m3), a 125% relative improvement in R2 compared to meteorological -variable-only predictions. Further application of the LFPM model enhanced prediction accuracy for all nine machine - learning methods, with XGBoost still leading (R2 = 0 . 873, RMSE = 8.17 μg/m3) .These ﬁndings conclusively demonstrate that integrating lagged feature variables signiﬁcantly enhances ozone prediction accuracy, offering stronger support for environmental monitoring and policy-formulation. KEYWORDS ozone, meteorological variables, pollutant variables, machine learning, prediction 1 Introduction With the accelerating pace of urbanization, air pollution has emerged as a critical global challenge, where ozone (O3) concentration dynamics have become a pivotal indicator of atmospheric quality (Ellingsen et al., 2008; Li et al., 2019) . As a secondary pollutant formed through complex photochemical processes, O3 levels exhibit strong dependencies on meteorological parameters, vehicular emissions, and industrial activities (Suciu et al., 2017; Wang et al., 2017; Tan et al., 2022) . While

FAQs

What data sources were used to predict ground-level ozone in Beijing?

The model used historical ozone and nitrogen dioxide (NO2) concentrations from the previous three hours as lagged inputs, combined with meteorological variables, and later added additional pollutant variables to improve performance.

Which machine-learning approach performed best overall?

XGBoost achieved the highest predictive accuracy among the evaluated methods, reaching R2=0.767 with pollutant additions and further improving to R2=0.873 when using the LFPM framework.

Why did adding lagged and pollutant variables improve prediction accuracy?

Including lagged feature variables captured temporal dependence of ozone and NO2, while adding pollutant variables improved modeling of nonlinear relationships between ozone, precursors, and meteorological drivers, leading to stronger accuracy across methods.

Aug 3, 2026

Uploaded by Rowan