Knowledge distillation for speed-acceleration distribution prediction at signalized intersections

Accurate emission estimation at signalized intersections requires the joint speed-acceleration, distribution, yet obtaining it has traditionally required trajectory data that are expensive and difficult to scale. This study proposes a knowledge distillation framework combining Gaussian Mixture Models and Ridge Regression to predict this distribution from 10 features available from standard traffic monitoring, addressing the data bottleneck limiting micro-level emission model deployment. A teacher model trained on 39 features from 35 million records at five Shanghai intersections transfers knowledge to a student model requiring standard inputs, achieving R-squared values of 0.81–0.99 across five distribution parameters for lane-specific entrance and exit models. The predicted distribution maps directly onto the operating mode bin fractions required by MOVES, enabling intersection-level emission estimation from routine traffic data. Entrance lanes are more infrastructure-sensitive while exit lanes are more temporally sensitive, with implications for intersection design and signal optimization.