A deep learning framework to generate synthetic mobility data

Synthetic datasets are useful when real-world data is limited or unavailable. They can be used in transport simulation models to predict travel behavior or estimate demand for transportation services. However, building these models requires large amounts of data. We propose a novel framework to generate a synthetic population with trip chains using a combination of generative adversarial network (GAN) with recurrent neural network (RNN). Our model is compared with other recent methods, such as Composite Travel Generative Adversarial Networks for Tabular and Sequential Population Synthesis (CTGAN) and shows improved results in predicting trip distributions. The model is evaluated using multiple assessment metrics to gauge its performance and accuracy.

I. Introduction

Census and household travel surveys are conducted and published every 5 or 10 years in most countries. They provide rich content about the demographic information of households and geospatial information on the daily trip routines these households have. The Census dataset includes socioeconomic characteristics of the population such as age, sex, and industry, while household surveys include variables related to the personal trip behaviors of the participants in addition to socioeconomic variables. Although Census data provides various demographic features of the households, the information about travel behavior is mostly not given or very limited. On the other hand, household travel surveys are more travel oriented and provides more details in relation to the travel features of the households.