저소음 포장도로 소음 예측모델별 성능 비교 및 평가

키워드

확장요약문

Road traffic noise is a major source of environmental noise in urban areas and can adversely affect public health and quality of life. In Seoul, road traffic noise accounts for more than 40% of noise-related complaints. Although noise barriers are widely used for mitigation, their effectiveness can be limited in densely built urban areas. Low-noise pavement has therefore emerged as an alternative approach that reduces noise at its source. This study compares the predictive performance of ensemble-based machine learning models and investigates changes in feature importance before and after low-noise pavement installation.

Noise measurements were conducted along the Inner Ring Road in Seoul, from Hongjecheon Overpass to Gajwa Station, where Stone Mastic Asphalt (SMA) pavement was installed. Noise levels were measured using a RION NL-52 sound level meter in accordance with Korean official measurement procedures. Hourly equivalent continuous sound levels were calculated, and traffic volume and vehicle speed data were obtained from the Seoul Transport Operation and Information Service (TOPIS). Three machine learning models—Extreme Gradient Boosting (XGBoost), Gradient Boosting Machine (GBM), and Random Forest (RF)—were evaluated using the coefficient of determination (R2), root mean squared error (RMSE), and mean absolute error (MAE). Feature importance was assessed using model-specific metrics.

The R2 values of GBM and RF decreased from 79.95% to 63.30% and from 75.04% to 55.08%, respectively. In contrast, XGBoost showed an increase from 65.40% to 77.43%. RMSE and MAE decreased for all three models after installation, indicating lower prediction errors. XGBoost showed improvements in both explanatory performance and prediction error metrics. For XGBoost, traffic volume importance decreased from 0.4805 to 0.4606, while vehicle speed importance increased from 0.5195 to 0.5394. Similar trends were observed for GBM and RF. These consistent changes suggest that vehicle speed became relatively more influential in noise prediction under the post-installation condition.