PREDICTING INTERSECTION QUEUE WITH NEURAL-NETWORK MODELS

To capture the complex nature of intersection queue dynamics, this study has explored the use of neural network models with data from extensive simulation experiments. The proposed models, although lacking in mathematical elegance, are capable of providing the acceptable prediction accuracy (more than 90%) at 3 time-steps ahead. As each time-step is as short as 3 s, the resulting information on queue evolution is sufficiently detailed for both responsive signal control and intersection operations. To accommodate the differences in available surveillance systems, this study has also investigated the most suitable neural network structure for each proposed queue model with extensive exploratory analyses.