A multi-task deep learning model for short-term taxi demand forecasting considering spatiotemporal dependences

Huimin Luo, Jianming Cai, Kunpeng Zhang, Ruihang Xie, Liang Zheng
2020 Journal of Traffic and Transportation Engineering (English ed. Online)  
A multi-task deep learning model for short-term taxi demand forecasting considering spatiotemporal dependences, Journal of Traffic and Transportation Engineering (English Edition), https://doi.9 Highlights 10 • The multi-task deep learning model is built to forecast the future change of the multiple tasks at the same 11 time. 12 • The nonlinear Granger causality analysis can capture the spatiotemporal correlations among the different 13 traffic zones. 14 • Bayesian optimization is promising to
more » ... ptimize the hyperparameters of deep learning models. 15 Abstract 16 Short-term taxi demand forecasting is of great importance to incentivize vacant cars moving from over-supply 17 regions to over-demand regions, which can minimize the wait time for passengers and drivers. With the 18 consideration of spatiotemporal dependences, this study proposes a multi-task deep learning (MTDL) model 19 to predict short-term taxi demand in multi-zone level. The nonlinear Granger causality test is applied to 20 explore the causality relationships among various traffic zones, and long short-term memory (LSTM) is used 21 as the core neural unit to construct the framework of the multi-task deep learning model. In addition, several 22 hyperparameter optimization methods (e.g., grid search, random search, Bayesian optimization, hyperopt) are 23 used to tune the model. Using the taxi trip data in New York City for validation, the multi-task deep learning 24 model considering spatiotemporal dependences (MTDL*) is compared with the single-task deep learning 25 model (STDL), the full-connected multi-task deep learning model (MTDL # ) and other benchmark algorithms 26 (such as LSTM, support vector machine (SVM) and k-nearest neighbors (k-NN)). The experiment results show that the proposed MTDL model is promising to predict short-term taxi demand in multi-zone level, the 1 nonlinear Granger causality analysis is able to capture the spatiotemporal correlations among various traffic 2 zones, and the Bayesian optimization is superior to the other three methods, which verified the feasibility and 3 adaptability of the proposed method. 5 Actually, traffic demand prediction is not only a time series problem, but also has spatial dependence 13 characteristics. In order to simultaneously capture the temporal and spatial correlations in the end-to-end 14 architecture, researchers have recently attempted to apply deep learning to traffic demand forecasting. For 15 example, Ke et al. (2017) employed a deep learning approach that fused convolutional techniques and the 16 long short-term memory (LSTM) network (FCL-net) to predict the passenger demand under an on-demand 17 ride service platform. Experimental results showed that the FCL-net had better predictive performance than 18 the classical time series prediction models and state-of-art machine learning algorithms. Liu and Chen (2017) 19 proposed an hourly passenger flow prediction model using deep learning methods, the experimental results of 20 which indicated that the proposed model can provide a more accurate and universal passenger flow prediction 21 model for different bus rapid transit (BRT) stations with different passenger flow profiles. As one type of traffic
doi:10.1016/j.jtte.2019.07.002 fatcat:2vgiq3zwhjelxm2bkel6dzwtz4