Predicting network flows from speeds using open data and transfer learning

Mahajan, Vishal; Cantelmo, Guido; Rothfeld, Raoul; Antoniou, Constantinos · 2023 · DOAJ

DOI: 10.1049/itr2.12305

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of traffic flow data scarcity, which hinders the calibration and validation of traffic simulation models. While traffic speed data is widely available from Floating Car Data (FCD) sources, traffic flow (volume) data is expensive to collect and limited to specific sensor locations. The authors propose an indirect estimation approach that predicts network-wide traffic flows using only speed data and exogenous features, without relying on time-lagged flow values. The study investigates two primary research questions: how accurately dynamic link flows can be estimated from open speed data, and under what conditions transfer learning can successfully generalize these models across different cities. The methodology employs Long Short-Term Memory (LSTM) deep learning models trained on real-world, publicly available data from Paris and Madrid. The authors fuse heterogeneous datasets, including traffic counts and speed traces, to create a comprehensive training set. The model utilizes static predictors (e.g., road geometry, speed limits) and dynamic exogenous predictors (e.g., current speed, time of day) to forecast future traffic flows. The experimental design involves training the model on Paris data and testing its performance on Madrid data to evaluate transfer learning capabilities. The authors also conduct sensitivity analyses on input-output sequence lengths and compare the performance of transferred models against those trained from scratch on limited target data. The results demonstrate that prevalent speed data, when combined with relevant geometric and contextual features, can reasonably forecast traffic flows. The study finds that transfer learning is particularly beneficial when data for the target task is minimal; in such scenarios, models pre-trained on source data outperform those trained from scratch. Success in transfer learning depends on the training set adequately capturing the flow-speed relationship and the target data having similar spatial-temporal characteristics to the source data. The LSTM models proved competitive with more complex graph-based architectures for this specific indirect estimation task. The significance of this work lies in its ability to scale traffic flow information from sparse sensor locations to entire networks using open data. By providing a method to generate flow estimates where direct measurements are unavailable, the approach aids in overcoming data scarcity challenges in transportation modeling. This facilitates better calibration of large-scale traffic simulation models and improves dynamic origin-destination demand estimation. The use of publicly available data ensures that the methodology is replicable and accessible for researchers and practitioners aiming to enhance traffic management systems without incurring high data collection costs.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• telematics crash prediction