Air pollution forecasting using advanced machine learning techniques and ensemble stacking in Delhi

Background: Air pollution, primarily due to air particulates like PM10 and PM2.5, causes several respiratory health problems. Accurate forecasting of particulate matter concentration is crucial for managing complex and nonlinear data, allowing timely interventions for early warning systems and air pollution control. The study aimed to develop reliable machine learning models for forecasting PM2.5 and PM10 concentrations, providing actionable insights for air quality management and public health. Methods: Despite challenges with unusual patterns and abrupt changes, the models achieved high accuracy, with R2 values exceeding 0.96 and low RMSE values. MLP outperformed the RF and XGB models for both PM2.5 and PM10 predictions. MLP-based stacking models further enhanced prediction accuracy, achieving the lowest RMSE and highest R2 values. For PM10, the weighted average approach provided better performance, striking an optimal balance between the different models’ contributions. Results: Despite unusual patterns and rapid jerks, our models had the highest R2 (>0.96) and lowest RMSE values. MLP outperformed the RF and XGB models for both pollutants. It improves the PM2.5 concentration predictions of stacking models, notably those using MLP as the meta-learner. MLP-based stacking yielded the lowest error values. The weighted average strategy improved the PM10 performance more than the stacking models and provided a better balance between the model contributions. Conclusion: Ensemble models achieved enhanced predictive accuracy by emphasizing the importance of selecting machine learning models and stacking methods based on air contaminants and data features, which is a crucial aspect of air quality management and public health.