Comparative Analysis of Machine Learning and Deep Learning Models for PM2.5 and PM10 Time Series Forecasting Using the SISANAPAS Web Platform

Abstract

Particulate matter (PM2.5 and PM10) pollution remains a significant environmental concern in Indonesia. This study employed the SISANAPAS web platform to compare machine learning (ML) and deep learning (DL) algorithms for PM2.5 and PM10 forecasting. Using 3-hourly data from Cibeureum (January-October 2024), which underwent comprehensive pre-processing (K-Nearest Neighbors imputation, Z-score outlier removal, Min-Max scaling, feature engineering), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Random Forest, and XGBoost models were evaluated. XGBoost Regression provided the most accurate forecasts, with an R² of 0.85, RMSE of 5.05 µg/m³, and MAE of 3.19 µg/m³ for PM2.5, and an R² of 0.85, RMSE of 9.71 µg/m³, and MAE of 6.55 µg/m³ for PM10. These results, significantly outperforming LSTM and GRU, highlight XGBoost's potential for reliable air quality prediction and demonstrate SISANAPAS as a valuable tool for environmental data analysis, crucial for informing public health and environmental policies.