Forecasting the Air Quality Index Using Machine Learning Models, Bayesian Optimization, and the Development of the S-GBR Model Incorporating Seasonal Variables

Air pollution is considered one of the most serious environmental and public health challenges in urban communities, and accurately forecasting the Air Quality Index (AQI) plays a crucial role in mitigating its negative impacts and supporting data-driven decision-making. Given the complexity and nonlinear nature of factors influencing air quality, the use of machine learning methods has attracted widespread attention in recent years. However, a review of previous studies reveals two major shortcomings: first, many models have been implemented based on default hyperparameter values, which has led to reduced accuracy and generalizability; second, temporal and seasonal components have often been overlooked, even though they play a decisive role in variations in air quality. To address these shortcomings, this study proposes a novel framework called the Seasonal Gradient Boosting Regressor (S-GBR). In this model, the Bayesian optimization search method is used for hyperparameter optimization, and the seasonal feature is incorporated as an input to the Gradient Boosting Regressor algorithm. In addition, baseline models such as Random Forest and XGBoost were also simulated and compared to determine the standing of the proposed model. Empirical findings show that the proposed model achieved a coefficient of determination of ۰.۹۶۸۶ and significantly reduced errors, performing almost as well as the most accurate baseline model (Random Forest with ۰.۹۷۹۶) while outperforming XGBoost. These results demonstrate that combining Bayesian optimization with the inclusion of seasonal components can raise prediction accuracy to the level of rich and complex datasets, even under limited data conditions. Such an achievement highlights the high potential of the proposed model for use in practical air quality monitoring and management. Air pollution is considered one of the most serious environmental and public health challenges in urban communities, and accurately forecasting the Air Quality Index (AQI) plays a crucial role in mitigating its negative impacts and supporting data-driven decision-making. Given the complexity and nonlinear nature of factors influencing air quality, the use of machine learning methods has attracted widespread attention in recent years. However, a review of previous studies reveals two major shortcomings: first, many models have been implemented based on default hyperparameter values, which has led to reduced accuracy and generalizability; second, temporal and seasonal components have often been overlooked, even though they play a decisive role in variations in air quality. To address these shortcomings, this study proposes a novel framework called the Seasonal Gradient Boosting Regressor (S-GBR). In this model, the Bayesian optimization search method is used for hyperparameter optimization, and the seasonal feature is incorporated as an input to the Gradient Boosting Regressor algorithm. In addition, baseline models such as Random Forest and XGBoost were also simulated and compared to determine the standing of the proposed model. Empirical findings show that the proposed model achieved a coefficient of determination of ۰.۹۶۸۶ and significantly reduced errors, performing almost as well as the most accurate baseline model (Random Forest with ۰.۹۷۹۶) while outperforming XGBoost. These results demonstrate that combining Bayesian optimization with the inclusion of seasonal components can raise prediction accuracy to the level of rich and complex datasets, even under limited data conditions. Such an achievement highlights the high potential of the proposed model for use in practical air quality monitoring and management.

کلیدواژه ها:

Air Quality Index forecasting ، machine learning ، hyperparameter optimization ، Bayesian search ، seasonal features ، S-GBR model