Ozone Forecasting: A Comparative Study of Ensemble Learning Methods under Varying Meteorological Factors

This study develops and evaluates machine learning-based predictive frameworks for ground-level ozone estimation using key meteorological factors, such as maximum and minimum temperature, rainfall, wind speed, and relative humidity. The research objective is to develop a machine learning-based framework for advancing predictive capability and to inform targeted air-quality management strategies in data-scarce environments.

We applied five machine learning (ML) algorithms, XGBoost (XGB), Random Forest (RF), Gradient Boosting (GB), Deep Neural Network (MLP), and Support Vector Regression (SVR), to develop ML models. The ML frameworks were systematically benchmarked to capture the nonlinear and interacting effects of weather variables on ozone dynamics. We studied the weather data from a publicly available source for an ozone-rich region in the northwestern part of Bangladesh.

We found that XGBoost consistently outperformed its counterparts among other models. It achieved higher accuracy, measured by the regression analysis, and robustness in tracing ozone concentration patterns across diverse atmospheric conditions. Beyond predictive performance, a feature importance analysis was conducted, with the temperature range (the difference between maximum and minimum temperatures), rainfall, wind speed, and relative humidity. We calculated that maximum temperature is the most influential factor governing both ozone accumulation and dispersion within the study region. This finding underscores the critical role of thermal variability in photochemical processes and boundary-layer mixing. In summary, the results highlight the efficacy of ensemble tree-based methods, particularly XGBoost, for operational ozone forecasting, while offering actionable insights into the physical mechanisms underlying pollutant behavior.