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

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