Construction and Implementation of an Enhanced WA-Stacking Ensemble Model for Air Pollutant Prediction

Currently, most individual machine learning models have limited effectiveness in predicting long-term air pollutant concentrations, especially when considering multiple influencing factors as inputs. The prediction results tend to fall into local optima and suffer from overfitting. To address these issues, this paper combines two types of influencing factors—air pollutant concentration datasets and meteorological factor datasets—to establish a Stacking ensemble prediction model. This model integrates three base learners: Random Forest, LightGBM, and Extreme Gradient Boosting (XGBoost). The construction mechanism of this ensemble model effectively consolidates the advantages of Random Forest, LightGBM, and XGBoost, enabling the input of multiple influencing factors without easily overfitting. It also enhances the model’s stability and predictive capability, improving its robustness and generalization performance. Additionally, LSTM (Long Short-Term Memory), as a time-series model, is incorporated by taking a weighted average of its predictions with those of the Stacking ensemble model, thereby constructing a WA-Stacking integrated prediction model. The overall results demonstrate that the WA-Stacking ensemble prediction model achieves more significant performance, making it suitable for research in the prediction and prevention of air pollutants.