GC-GRU-N for Traffic Prediction using Loop Detector Data

TL;DR

This paper introduces GC-GRU-N, a graph convolutional gated recurrent unit network for traffic prediction that effectively captures spatiotemporal dependencies, achieving competitive accuracy with significantly faster inference than transformers.

Contribution

The paper proposes a novel GC-GRU-N model that combines graph convolution with gated recurrent units for improved traffic prediction performance.

Findings

GC-GRU-N ranks second among benchmark models.

It has six times faster inference than transformers.

Achieves comparable accuracy to transformers on key metrics.

Abstract

Because traffic characteristics display stochastic nonlinear spatiotemporal dependencies, traffic prediction is a challenging task. In this paper develop a graph convolution gated recurrent unit (GC GRU N) network to extract the essential Spatio temporal features. we use Seattle loop detector data aggregated over 15 minutes and reframe the problem through space and time. The model performance is compared o benchmark models; Historical Average, Long Short Term Memory (LSTM), and Transformers. The proposed model ranked second with the fastest inference time and a very close performance to first place (Transformers). Our model also achieves a running time that is six times faster than transformers. Finally, we present a comparative study of our model and the available benchmarks using metrics such as training time, inference time, MAPE, MAE and RMSE. Spatial and temporal aspects are also analyzed for each of the trained models.