Vehicle Travel Time Estimation Using Sequence Prediction

Gündüz, Gültekin; Acarman, Tankut · 2020 · Crossref

DOI: 10.7307/ptt.v32i1.3008

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 short-to-medium-term vehicle travel time and traffic speed prediction, a critical component of Intelligent Transportation Systems. The authors propose a region-based sequence prediction method that leverages Recurrent Neural Networks (RNNs) to capture non-linear spatio-temporal relationships in traffic data. The motivation stems from the limitations of traditional parametric and non-parametric approaches, which often struggle with the complex dynamics of urban traffic, including varying congestion levels and noise in Global Navigation Satellite System (GNSS) data. The study utilizes Floating Car Data (FCD) collected from 8,317 passenger vehicles in Istanbul, Turkey, over a 34-day period in January and February 2018. The data, comprising latitude-longitude coordinates, timestamps, and vehicle velocities, were analyzed across twelve distinct districts selected for their varied road characteristics, such as speed limits, curvatures, and urban canyon effects. The methodology involves breaking down entire vehicle trips into smaller "sub-trips" defined by consecutive GNSS measurements with an elapsed time of at most 15 seconds. An unsupervised agglomerative clustering algorithm extracts directed links between road segments. For each district, two RNNs with Long Short-term Memory (LSTM) cells are trained: one to predict the travel time ($\Delta t$) of a sub-trip and another to predict the ending speed ($v_{t+1}$). The input sequences include meta-information such as the day of the week, district coordinates, and start/end velocities. The models were trained using 25% of the data, validated with 25%, and tested on the remaining 50%, with a batch size of 32 and 50 epochs. The results demonstrate that the proposed LSTM-based sequence prediction method outperforms both Multilayer Perceptron (MLP) function mapping and linear velocity change models. Across all twelve districts, the RNN approach achieved lower Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) for both travel time and speed estimations. For instance, in District D1, the RNN achieved an RMSE of 1.384 seconds for travel time, compared to 2.076 seconds for the MLP and 1.436 seconds for the linear model. Similarly, for speed estimation in District D1, the RNN recorded an RMSE of 5.205 km/h, significantly lower than the MLP’s 7.633 km/h. The analysis indicates that while some districts exhibited higher prediction error variance, the RNNs successfully captured the step-like changes in traffic dynamics. The significance of this work lies in its ability to provide continuous arterial speed information and accurate travel time estimates by learning district-specific non-linear dynamics. By utilizing sequence prediction with LSTM cells, the method effectively handles the irregularities and noise inherent in GNSS data, offering a robust alternative to stationary sensor-based systems. This approach enhances the precision of Intelligent Transportation Systems, enabling better traffic management and route planning in metropolitan areas.

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