Deep-Learning Architectures to Forecast Bus Ridership at the Stop and Stop-To-Stop Levels for Dense and Crowded Bus Networks

Abstract

The conventional transit assignment models that depend on either probabilistic or deterministic theory have failed to accurately estimate rider demand for dense and crowded bus transit networks. It is well known that the existing models are so blunt that they cannot accommodate the impact of miscellaneous changes in activity and transportation systems on bus demand. Recently, artificial neural networks (ANNs) have been refocused after two monumental breakthroughs: Big-data and a novel pre-training method. A deep-learning model, which simply represents an ANN with multiple hidden layers, has had a great success in recognizing images, human voices, and handwritten texts. The present study adopted a deep-learning model to forecast bus ridership at the stop and stop-to-stop levels. While the stop-level model, which had insufficient training data, suffered from an overfitting of the data, the stop-to-stop-level model showed good performance both in training and testing. The success of the latter model is owed to a larger sample size compared with the former model. This represents the first meaningful attempt to apply a data-driven approach to forecasting transportation demand.