Comparison of Time Series Regression, Support Vector Regression, Hybrid, and Ensemble Method to Forecast PM2.5

Introduction: PM2.5 pollution poses significant health risks, particularly in Jakarta, where levels often exceed safety thresholds. Accurate forecasting models are essential for air quality management and mitigation strategies. Methods: This study compares four forecasting models: Time Series Regression (TSR), Support Vector Regression (SVR), a hybrid TSR-SVR model, and an ensemble approach. The dataset consists of 9,119 hourly PM2.5 observations from January 1, 2023, to January 15, 2024. Missing values were imputed using historical hourly trends. Model performance was evaluated using Root Mean Squared Error (RMSE). Results: The hybrid TSR-SVR model achieved the lowest RMSE (6.829), outperforming TSR (7.595), SVR (7.477), and the ensemble method (7.486). The hybrid approach effectively captures both linear and nonlinear patterns in PM2.5 fluctuations, making it the most accurate model. Conclusions: Integrating statistical and machine learning models improves PM2.5 forecasting accuracy, aiding policymakers in pollution control efforts. Future studies should explore additional external factors to enhance prediction performance.