An Intelligent Traffic Flow Progression Model for Predictive Control Applications: A Proposed Approach for Making Short-term Traffic Flow Predictions, Using a Recurrent, Multi-task Learning Neural Network Model, to Transform Traffic Light Controllers from Adaptive to Predictive with Minimal Hardware Changes

This study explores the possibility of developing a short-term traffic flow prediction model that can be used to convert installed adaptive controllers to predictive controllers with minimal hardware changes. By using the prediction model’s outputs to virtually trigger vehicle loop detectors, the outputs of an adaptive controller can be extracted in advance of actual vehicle arrivals. This will enable service providers to send out time to green/red (T2G/R) information or green light optimal speed advice (GLOSA), which are driver assistance use cases that aim to efficiently guide vehicles through intersections, in anticipation of known upcoming signal states. The main requirements for developing the prediction model are that it should be scalable to different intersection configurations, adaptable to different traffic conditions and should encompass the nonlinearity of traffic flow behavior. Since it fits these criteria, the developed model is a multi-task learning recurrent neural network with exogenous inputs (NARX), which is designed to match the traffic flow simulation abilities of a well adopted analytical traffic flow progression model. Both models were tested for a corridor in Delft that experiences almost consistent free flow conditions, and on an intersection in Haarlem with varying traffic conditions. For both case studies the neural network outperformed the analytical model. Most notably, it was better adaptable to long queues at intersections, had lower average error values, made fewer large errors, and better recognized the effect of a source/sink. When interfaced with an adaptive controller to test the predictive control methodology proposed, the superiority of the designed neural network model over the analytical model became more prominent. At no point did the neural network’s prediction errors result in queue spill-back, which was not true for the analytical model. However, the overall accuracy of the NARX model was still not yet satisfactory enough for practical application (especially for highly under-saturated traffic conditions) without the use of corrective measures. Nonetheless, due to its significant superiority over the widely used analytical model, with regards to both accuracy and adaptability, this model can be considered as a new starting point for traffic flow progression modeling.Transport, Infrastructure and Logistic