Deep Learning based Prediction of Vehicular Speed, Flow and Traffic Congestion using Real World Data

Accurate traffic congestion and flow prediction plays a crucial role in designing intelligent transport system (ITS) for vehicles in heterogeneous mobile networks. Traffic speed, flow and congestion prediction is a challenging task if limited attributes are used as input to predictive models or the prediction is done too late to take any appropriate action. Although several deep learning and machine learning approaches have been deployed to predict different traffic conditions, these approaches are based on the traffic observations at target location or its adjacent regions. Complex road networks, large data sets with multiple traffic features and analysis of spatial temporal dependencies of traffic data are not yet exploited in detail. In this study, a workflow consists of data acquisition, analysis, and exploitation of real-world time series traffic data is developed. We study number of data driven models namely, the long short-term memory (LSTM), gated recurrent unit (GRU) and hybrid CNN-LSTM model to predict traffic and compare it with the linear regression (LR) model. The deep learning models are capable of handling both spatial and temporal dependencies as well as the sudden changes in traffic speed and flow predicting the vehicular traffic accurately over long periods. The models are trained using the six months and three months of traffic flow, speed and occupancy data provided by the California Department of Transportation (Caltrans). The deep learning models outperform traditional linear regression method. Moreover, an architectural study of deep learning models is carried out for traffic prediction to discover the best structure for traffic congestion prediction. The performance of the models is evaluated using MSE and MAE metrics. It is concluded that the performance of deep learning models vary with the amount of historical data and sliding window size used for the traffic prediction. It is observed that a complex hybrid model like CNN-LSTM is not required for accurate prediction of the spatial and temporal tendencies when the training period is longer (six months). The GRU with much simpler architecture does not need to store road network and temporal information for long term in its memory, outperforms the LSTM and CNN- LSTM models. However, as the data size is reduced CNN-LSTM model out- performs the LSTM and GRU models for traffic congestion prediction.