Immediate Velocity Prediction of a Mixed-Traffic Flow in Urban Road Networks with Spatial-Temporal Correlation Analysis

Enhancing the precision of short-term transportation flow forecasting is vital for traffic engineering applications. By selecting appropriate characteristics from collected data to strengthen the prediction model, the performance and the exactness of predicted velocity would be improved considerably. From the observation of speed tracking figures in a complex urban road network, the velocity at a specific road segment is intimately affected by immediate past traffic conditions of its adjacent links. In this paper, we evaluate various hypotheses on multiple datasets to determine the impact of the spatial besides temporal aspect on the concluding prediction. A recurrent neural networks variation - Long Short-Term Memory - is experimented with thoroughly tuning on the hyper-parameters. The result shows that the more spatial knowledge we add, the more precise the predicting velocity is, especially on crowded intersections. Moreover, the traffic speed does not correlate with extensive historical data and the prediction accuracy is better with the aggregated velocity of each road segment.