A Hybrid DNN Model for Travel Time Estimation from Spatio-Temporal Features

Abstract

Due to recent advances in the Vehicular Internet of Things (VIoT), a large volume of traffic trajectory data has been generated. The trajectory data is highly unstructured and pre-processing it is a very cumbersome task, due to the complexity of the traffic data. However, the accuracy of traffic flow learning models depends on the quantity and quality of preprocessed data. Hence, there is a significant gap between the size and quality of benchmarked traffic datasets and the respective learning models. Additionally, generating a custom traffic dataset with required feature points in a constrained environment is very difficult. This research aims to harness the power of the deep learning hybrid model with datasets that have fewer feature points. Therefore, a hybrid deep learning model that extracts the optimal feature points from the existing dataset using a stacked autoencoder is presented. Handcrafted feature points are fed into the hybrid deep neural network to predict the travel path and travel time between two geographic points. The chengdu1 and chengdu2 standard reference datasets are used to realize our hypothesis of the evolution of a hybrid deep neural network with minimal feature points. The hybrid model includes the graph neural networks (GNN) and the residual networks (ResNet) preceded by the stacked autoencoder (SAE). This hybrid model simultaneously learns the temporal and spatial characteristics of the traffic data. Temporal feature points are optimally reduced using Stacked Autoencoder to improve the accuracy of the deep neural network. The proposed GNN + Resnet model performance was compared to models in the literature using root mean square error (RMSE) loss, mean absolute error (MAE) and mean absolute percentile error (MAPE). The proposed model was found to perform better by improving the travel time prediction loss on chengdu1 and chengdu2 datasets. An in-depth comprehension of the proposed GNN + Resnet model for predicting travel time during peak and off-peak periods is also presented. The model’s RMSE loss was improved up to 22.59% for peak hours traffic data and up to 11.05% for off-peak hours traffic data in the chengdu1 dataset.

A Graph Neural Network (GNN)-Based Approach for Real-Time Estimation of Traffic Speed in Sustainable Smart Cities

“…A real-time traffic-prediction model based on graph convolutional networks (GCNs) is presented in this paper. The model uses GCNs to represent geographical relationships in the traffic data and integrates the road network's graph structure [35]. In a study revealed in [36], a traffic-state estimation and prediction model that combines Recurrent Neural Networks (RNNs) and graph convolutional networks (GCNs) is proposed.…”