Comparative Analysis of Machine Learning Models for Multi-Horizon PM2.5 Forecasting

Accurate forecasting of particulate matter (PM2.5) concentrations is critical for public health management and environmental policy-making. This study presents a comprehensive comparison of six machine learning models—Linear Regression, Support Vector Regression (SVR), Random Forest (RF), Gradient Boosting Decision Trees (GBDT), Multi-Layer Perceptron (MLP), and Long Short-Term Memory (LSTM)—for multi-horizon PM2.5 prediction. Using hourly air quality data from 11 cities in Zhejiang Province, China (January-February 2024), we evaluate model performance across three forecast horizons: 1-hour, 6-hour, and 24-hour ahead predictions. Our results demonstrate that model performance varies significantly with forecast horizon. For short-term (1-hour) predictions, Linear Regression achieves the best performance (RMSE=10.682, R²=0.901), suggesting near-linear temporal dynamics. For longer horizons (24-hour), ensemble tree-based models outperform others, with GBDT achieving RMSE=24.264 and R²=0.467. Surprisingly, deep learning approaches (LSTM) underperform traditional machine learning methods, particularly for long-term forecasting. Feature importance analysis reveals that the most recent PM2.5 value (lag-1) accounts for 47.8% of predictive power, while Air Quality Index contributes 42.3%, highlighting the dominance of temporal autocorrelation in PM2.5 dynamics.