City-wide traffic flow prediction is crucial for Intelligent Transportation Systems (ITS), providing the foundation for effective traffic management and pollution monitoring in transportation. Despite advances in traffic modelling with the help of simulation, statistics, and machine learning, predicting traffic flow for entire cities in practice remains a challenge due to the inherent spatiotemporal complexity of the problem and the inconsistent availability of input data. This research investigates the viability of cross-city transfer learning for city-wide traffic flow prediction and tests whether models trained on data-rich networks can be effectively transferred to data-scarce cities. Recent developments in deep neural networks for spatiotemporal modelling, transfer learning for generalising pre-trained models to domains with limited data, and opportunistic data sources to complement sparse traffic counting sensors have created new opportunities for research on model generalisation across road links and networks. Therefore, we propose a three-step workflow consisting of (i) data preparation, (ii) training a deep neural network for city-wide traffic flow prediction in a data-rich urban network, and (iii) transferring the pre-trained model for network-wide predictions on a second city with minimal historical data. We demonstrate and discuss the performance and limitations of the proposed workflow for estimating traffic flows in Copenhagen, a data-scarce city, using a model pre-trained in Paris, a data-rich network. Our preliminary results reveal valuable insights for the practical deployment of machine learning models in real urban transportation systems.
Gualda et al. (Thu,) studied this question.