Deep Learning System for Travel Speed Predictions on Multiple Arterial Road Segments

Abstract

Accurate travel speed prediction is a critical tool for incidence response management. The complex dynamics of transport systems render model-based prediction extremely challenging. However, the large amounts of available vehicle speed data contain the complex interdependencies of the target travel speed; the data itself can be used to generate accurate predictions using deep learning methods. In this work, a deep learning methodology involving feature generation, model development, and model deployment is presented. The authors demonstrate the high performance of deep learning methods (relative to more traditional benchmarks) in predicting travel speeds from 5–30 min in advance, for a challenging arterial road network. In this study, different deep learning architectures that exploit both spatial and temporal information for several time frames are compared and analyzed. Finally, the authors demonstrate the integration of their deep learning method into a visualization system that can be directly applied for vehicle speed prediction in real time. The model-selection analysis and data-to-visualization framework in this manuscript provide a step towards decision support for incident management; for practical implementation, the predictive power of deep learning models under incident conditions should continue to be investigated and improved.

Deep Learning Framework for Freeway Speed Prediction in Adverse Weather

“…Dropout regularization is applied to mitigate the overfitting problems that traditional neural networks have. Recent studies used dropout regularization for enhancing DL testing results ( 43 , 51 , 52 ). When dropout regulation is applied at a hidden neural network layer, a random thinned neural network layer is produced from the original layer.…”

“…The 1D CNN–LSTM model has been found to outperform others (e.g., ARIMA, linear regression, Gaussian process, k-nearest neighbor, LSTM) in predicting traffic speed ( 43 ). However, a model with the LSTM structure is not suitable for nonrecurrent traffic conditions data, comprising greater variation of speed between time steps, which does not follow a specific pattern ( 36 ).…”

“…The 1D CNN model was found to outperform others as it yielded the lowest RMSE and MAE for predicting speed in all types of adverse weather. Although CNN–LSTM was found to be optimal for predicting traffic speed under prevailing conditions ( 43 ), the results indicated that the LSTM and CNN–LSTM structures were not suitable for predicting traffic speed under adverse weather conditions because of their high RMSE and MAE values.…”

Motorway Traffic Flow Prediction using Advanced Deep Learning

“…More recently, recurrent neural networks (RNNs) -such as the long short-term memory model (LSTM) -have been designed to learn from sequences of data, and to capture long-term temporal patterns. LSTM was applied to traffic flow data [12], either alone of jointly with convolutional neuronal networks (CNN) [13], [14] in an attempt to capture the spatial road network information. Among the difficulties of deploying such models are their often complicated structure, the choice of parameters (such as the number of neurons or the non-linear functions), and the fact that neural networks have been long time regarded as "blackboxes" [15].…”