Modeling and Forecasting Individual On-Demand Upcoming Trips

This research is motivated by the emerging on-demand mobility service as a complementary supply to public transport to meet flexible needs of commuters. Meanwhile, it helps to mitigate traffic congestion and disruption of public transport business due to overload of the travel demand. Travel time is an essential determinant for planning individual upcoming trips. To this end, this research aims to identify relevant data driven indicators affecting the travel time of individual on-demand trips, develop deep learning models to predict the travel time as well as identify the computational costs. The research strategy adopted is the exploratory case study of Chengdu city, China. Large data sets (50 Gigabytes) of commuter on-demand ride requests, which were collected from the digital platform of a mobility service provider Didi Chuxing in China, resulted in using a convenience sampling approach and a quantitative research analysis method to achieve the research objectives. Artificial Neural Network (ANN), Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM) are developed, evaluated, and compared. The findings with ground truth data indicate that model inputs, such as the departure time, travel distance, traffic zones, and most importantly, travel speed, are highly relevant indicators influencing the travel time of individual on-demand trips. The LSTM model outperforms other models in term of accuracy of travel time prediction. Due to extremely large data sets, all neural network models require significant amount of computation time for model training.