Studies from 2014 indicate that traffic congestions cause a loss of 25 billion Euros per year for logistics companies in Germany. Taking into account the impact on climate change through fuel waste, the frustration and stress of drivers and customers, it affirms the importance of a continuous traffic flow.
Especially, considering the fast urban growth and the increasing delivery traffic, a reliable traffic congestion prediction system can heavily support traffic management and telematics.
For predicting traffic congestion, in the first step a data set has to be generated by monitoring the traffic volume and recording the traffic flow. The major challenge is to combine this time series information with sufficient spatial information and additional relevant features, such as construction sites, weather conditions or failure of public transport, in order to incorporate all possible influencing factors. This requires many different data sets, which may not be easily accessible and may not have the same format and terminology.
A machine learning model can be trained to recognize patterns in time series data depending on the time of day, holidays, other time-related events and the other additional features. The trained model should then be able predict congestions on unseen data and adapting the weights of the different input data sources and conditions automatically.
To process the spatial features of the traffic flow data, where the propagation of flow in, out and along the motorway are represented, a fully-connected convolutional neural network (CNN) might be used. Since CNNs are commonly used for evaluating static image data, they are not capable of modelling long-term dependencies of temporal features, e.g. accidents or weather conditions. For this, the state-of-the-art method are long-short-term memory network (LSTMs).
The combination of both techniques enables the traffic flow prediction based on temporal and spatial traffic information.