Spatio-Temporal Traffic Flow Forecasting on a City-Wide Sensor Network

Intelligent transportation systems (ITS) all around the world are collecting and processing huge amounts of data from numerous sensors to generate a ground truth of urban traffic. Such data has set the foundation of traffic theory, planning and simulation to create rule-based systems but it can also be very useful for time-series analysis to predict future traffic flow. Still, the acceptance for data-driven forecasting is quiet low in productive systems of the public sector. Without enough probe data from floating cars (FCD) ITS owners feel unable to reach an accuracy like private telecommunication or car manufacturing companies. On the other hand, investigating into FCD requires a thoughtful treatment of user privacy and a close look on data quality which can also be very time consuming. With this paper we prove that a modern deep learning framework is capable to operate on city-wide sensor data and produces very good results with even simple artificial neural networks (ANN). In order to forecast space-time traffic dynamics we are testing a Feed Forward Neural Network (FFNN) with different geotemporal constraints and can show where and when they have a positive but also a negative effect on the prediction accuracy.