Predicting network flows from speeds using open data and transfer learning

Traffic flow/volume data are commonly used to calibrate and validate traffic simulation models. However, these data are generally obtained from stationary sensors (e.g. loop detectors), which are expensive to install and maintain and cover a small number of locations in the transport network. On the other hand, Floating Car Data (FCD) are readily available at the network level, usually from a sample of vehicles. We present an indirect traffic flow estimation approach using transfer learning to address the traffic flow data scarcity and model generalization across cities. Using two cities (Paris and Madrid) as study areas, we demonstrate the indirect estimation using only exogenous features for flow prediction, mirroring limited predictive features without past link flows. Subsequently, we use the model pre-trained on data from Paris city and test on data from Madrid city, and investigate the scenarios for successful transfer learning. Overall, the training set must adequately capture the flow-speed relationship for successful indirect flow estimation. Transfer learning is beneficial when the data for the target task is minimal, in which case transferred models outperform newly trained models from scratch. Using real-world and publicly available data, our approach and models can help scale a smaller traffic flow dataset to a larger sample across cities.