TPE-LCE-SHAP: A Hybrid Framework for Assessing Vehicle-Related PM2.5 Concentrations

Abstract

This study proposes a novel hybrid approach for estimating and analyzing vehicle-related PM2.5 concentrations. The framework integrates the Local Cascade Ensemble (LCE) model, optimized using the Tree-structured Parzen Estimator (TPE) strategy, with SHapley Additive exPlanations (SHAP) to enhance interpretability. It utilizes datasets comprising air quality, meteorological, and traffic data collected from strategically placed sensors along the Nairobi Expressway. Key parameters include hourly traffic volume, average vehicle speed, humidity, wind speed, and temperature. The TPE-tuned LCE model outperformed benchmark algorithms including Random Forest (RF), Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), Adaptive Boosting (AdaBoost), and Multiple Linear Regression (MLR) achieved the lowest Mean Absolute Error (MAE) of 1.94, Mean Squared Error (MSE) of 21.50, Root Mean Squared Error (RMSE) of 4.64, Residual Standard Ratio (RSR) of 0.38, and the highest Coefficient of Determination (R2) of 0.87. SHAP analysis of TPE-tuned LCE model identified location, humidity, and wind speed as the most influential predictors of PM2.5 levels. This hybrid framework delivers robust predictive accuracy and actionable insights, making it a valuable tool for effective environmental management and policy making.

Keywords

• Nairobi
• air quality
• PM2.5
• machine learning
• SHAP