Machine learning framework for predicting urban road speed profiles and uncertainty

Road speed serves as a vital indicator for traffic managers and road designers due to its correlation with congestion, safety, and pollutant emissions. Traditional speed models focus mainly on predicting the 85 th percentile of road speed distribution under free-flow conditions, primarily in rural areas, often overlooking speed variability. This study proposes a machine learning-based framework to predict real conditions on urban road segments and construct prediction intervals for assessing uncertainty. Utilizing a dataset of 650 road segments, this work covers various roadway configurations. Compared to existing models, the proposed framework demonstrates superior performance in accuracy and coverage metrics. It achieves an explained variance of R 2 = 74.30 % , with RMSE = 6.89 km / h and MAPE = 19.93 % . The prediction interval boasts a coverage level of 90.85%, exceeding the nominal value of 90%, and exhibits the lowest mean amplitude, enhancing its informativeness. Key variables influencing speed profile prediction include time stamps (hour and day), road classification, boundary conditions, density of speed bumps, and median strip presence. Notably, a signalized intersection reduces speed by approximately 27%, while the presence of a median increases it by about 6.1%.