Disaggregate Short-Term Location Prediction Based on Recurrent Neural Network and an Agent-Based Platform. Issue 8 (August 2019)

• Record Type: Journal Article
• Authors: Dong, Yanjie; Polak, John; Sivakumar, Aruna; Guo, Fangce
• Is Part Of: Transportation research record. Volume 2673:Issue 8(2019)
• Page Start: 657
• Page End: 668
• Publication Date: 2019-08
• Subjects: Transportation -- Periodicals; Roads; Transport -- Périodiques; Routes -- Périodiques; Routes -- Conception et construction -- Périodiques; Roads; Transportation; 388.05
• DOI: 10.1177/0361198119845880
• Languages: English
• ISSNs: 0361-1981

Abstract

With the growing popularity of mobile and sensory devices, there has been a strong research interest in short-term disaggregate-level location prediction. Such predictive models have huge application potential in several sectors to change and improve people's daily life and experience. Existing methods in this research stream have mainly focused on the prediction of sequence of location, with valuable temporal information overlooked. In addition, data limitations have constrained the development and understanding from different algorithms. In this paper, the authors propose a recurrent neural network-based method (RNN and LSTM, long short-term memory) for the next and future location prediction. This model predicts the sequence in time, thus it can predict both when and where an individual will be in the future and the duration of the stay at each location. The predictive model is developed based on an agent-based simulation platform that can produce realistic spatial-temporal trajectory data at the individual level. Analysis of the simulated data has shown that RNN and LSTM are capable of predicting future locations with better results than other comparative methods, especially for agents with high location variability. Online prediction with true location information fed into the model later in the day would greatly improve the predicted results. However, significant variations can be observed at the zonal level, with all methods performing much better on frequently visited locations than less visited locations or irregular visits.