REAL TIME SHORT TERM TRAFFIC PREDICTION FOR A REGIONAL MIDAS NETWORK BASED ON A NEURAL NETWORK PREDICTION ALGORITHM



REAL TIME SHORT TERM TRAFFIC PREDICTION FOR A REGIONAL MIDAS NETWORK BASED ON A NEURAL NETWORK PREDICTION ALGORITHM

Authors

Abraham Narh, Transport Systems Catapult, Carl Goves, Transport Systems Catapult, Paul Bate, Transport Systems Catapult

Description

In this paper, we further explore the scalability of the algorithms by deploying to an expanded regional network that includes over 500 loop sensors on key parts of the strategic road network in England.

Abstract

Previously, we presented results of the potential transferability of neural network short-term traffic prediction to three geographical areas (Bristol, Manchester and Southampton) of the UK motorway network. In this paper, we further explore the scalability of the algorithms by deploying to an expanded regional network that includes over 500 loop sensors on key parts of the strategic road network in England. Considering a 15-minute forecast window, the results indicate the neural network predicts accuracy levels of 83-85% (matched transitions) and 49-50% (transition accuracy) for a regional area network. The transition accuracy implies that the neural network would correctly forecast 1 of every 2 predictions when a traffic state changes.

The results demonstrate that adopting an expanded geographical network model, rather than the local area network used in the previous work, gives a benefit of circa 3% in transition accuracy. The main challenges for such massive deployment of sensor data is the computational run time requirements. By investigating options for accelerating the training time of the models, we prove that deploying the algorithms on a GPU (graphics processing unit) enabled computer could potentially improve the runtime by 2-5 times than a stand-alone CPU (central processing unit).

Finally, we offer lessons learnt in the selection of the optimal network structure for a Neural Network prediction algorithm. For this analysis, we developed 91 scenarios based on the number of hidden nodes (between 100 and 5000) and amount of training and validation datasets (between 2 and 50 weeks) to assess the effect of different model configurations on the prediction accuracy. By considering the runtime requirements versus the prediction accuracy based on the different configurations, we conclude that a complicated structure (with large number of hidden nodes and training and validation data) even though may give a high level of accuracy, this increase in accuracy is marginal and not necessarily worth the extended run times needed to train the model. We found that a simple network structure can give up to 85% runtime savings whilst achieving equally a high level of accuracy within an error margin of circa -1% (compared to a sophisticated structure).

Publisher

Association for European Transport