A residual spatio-temporal architecture for travel demand forecasting. (June 2020)

• Record Type: Journal Article
• Authors: Guo, Ge; Zhang, Tianqi
• Is Part Of: Transportation research. Volume 115(2020)
• Publication Date: 2020-06
• Subjects: Demand forecasting -- CNNs -- CE-LSTMs -- Residual connection; Transportation -- Periodicals; Transportation -- Technological innovations -- Periodicals; 388.011
• DOI: 10.1016/j.trc.2020.102639
• Languages: English
• ISSNs: 0968-090X

Abstract

Highlights: An improved LSTM, which is the fusion of convolutional technique and the traditional LSTM cell in a effective way, termed CE-LSTM, is proposed in this paper. A novel architecture named RSTN is constructed, which is an end-to-end trainable model that can capture the exogenous dependences and spatio-temporal correlations of travel demand adequately. A residual connection fashion is introduced to RSTN, yielding an easy-training approach by reformulating a traditional prediction problem as a learning residual function with regard to the travel density in each time interval. A novel representation of the historical travel demand is proposed in this paper. It enables deeper and more logical representation of historical data and reflects the more detailed dynamics of travel requests. Abstract: This paper proposes a deep architecture called residual spatio-temporal network (RSTN) for short-term travel demand forecasting. It comprises fully convolutional neural networks (FCNs) and a hybrid module consisting of an extended Conv-LSTM (CE-LSTM) that can achieve trade-off of convolutional operation and LSTM cells by tuning the hyperparameters of Conv-LSTM, convolutional neural networks (CNNs) and traditional LSTM. These modules are combined via residual connections to capture the spatial, temporal and extraneous dependencies of travel demand. The end-to-end trainable RSTN redefines the traditional prediction problem as a learning residual function with regard to the travel density in each time interval. Further more, a dynamic request vector (DRV)-based data representation scheme is presented, which catches the intrinsic characteristics and variation of the trend, to improve the performance of forecasting. Simulations with two real-word data sets show that the proposed method outperforms the existing forecasting algorithms, reducing the root mean square error (RMSE) by up to 17.87%.