Citation

Saroj AJ, Guin A, Hunter M. J. Big Data Anal. Transp. 2021; 3(2): 95-108.

Copyright

(Copyright © 2021, Holtzbrinck Springer Nature Publishing Group)

DOI

10.1007/s42421-021-00043-2

PMID

unavailable

Abstract

This research investigates the performance of deep learning to perform traffic data imputation using Long Short Term Memory (LSTM) Recurrent Neural Network (RNN) layers. Investigation of real-time traffic volume data received from a connected corridor revealed the presence of intermittent data gaps. Such data gaps in data streams could impact applications that utilize connected corridor data. To improve the utility of the connected corridor real-time data streams a deep learning algorithm that can use historic and current high frequency data to learn and provide accurate estimations for imputation is realized. In this study traffic volume data streams are received in 6-min aggregate bins from the corridor detectors. Univariate time series models based on only the given detector and multivariate time series models based on the given detector and a similar detector are trained on LSTM RNN layers using current and historic data. To investigate the performance of these models in imputing missing data gaps two experiments are conducted. The first experiment investigates the performance of models to impute consecutive missing data under typical traffic conditions.

RESULTS indicate comparable performance of the multivariate and univariate models for shorter consecutive missing data gap imputations, while for longer consecutive missing data gaps, multivariate outperforms univariate for several cases. The second experiment compared the models performance under atypical traffic conditions.

RESULTS indicate improved performance of the multivariate over univariate models, further demonstrating the potential advantages of using recent information from other similar detectors in a multivariate model, under both typical and atypical conditions.

Language: en

Keywords

Connected corridor data; Deep learning; Long short term memory (LSTM); Recurrent neural networks (RNN); Time-series prediction; Traffic data imputations