Travel Demand Forecasting At Regional Level Considering Spatiotemporal Dependencies

Travel demand forecasting plays an important role in coordinating the balance of travel demand and supply in the urban public transportation system.The urban road network naturally divides the city into different regions,and accurate prediction of travel demand in different regions can help taxi operators and ride-hailing platforms preallocate idle vehicle resources,thereby increasing vehicle utilization and reducing passenger waiting time.Due to the differences in geographic location and function distribution,the distribution patterns of travel demand in different regions are obviously different,which poses a challenge to the capture of spatiotemporal dependences in travel demand forecasting.This thesis proposes a multi-task dynamic spatiotemporal graph attention network to explore the complex spatiotemporal correlations between different regions,and finish demand prediction of multiple regions based on multi-task learning technology.The main contents of this thesis are as follows:(1)Determine the region division for travel demand forecasting.First,regular grid division schemes and irregular region division schemes under three different scales are designed.Then,suitable travel demand forecasting models based on the characteristics of the data structure are established.By comparing and analyzing the prediction results based on the regular grid and irregular regions,the irregular region is taken as the final zoning scheme for travel demand forecasting.(2)Establish a multi-task dynamic spatiotemporal graph attention network to predict travel demand in different regions.First,encode the temporal dependence according to the key frame extraction method,and encode the spatial dependence according to the adjacent spatiotemporal graph and the far spatiotemporal graph.Then,the spatiotemporal learning component is constructed based on the recurrent graph attention network and graph learning mechanism to extract spatiotemporal dependencies.Finally,the community detection algorithm is used to divide all regions into different sets,and the demand prediction of the regions in each set is treated as a learning task,and a multi-task learning component is constructed to finish the prediction.The proposed model is verified based on two real-world datasets,and the effectiveness of each module is verified through various ablation experiments.Experimental results indicate that the proposed multi-task dynamic spatiotemporal graph attention network is superior to traditional time series prediction methods,classic machine learning methods and deep learning methods,and advanced spatiotemporal graph neural network methods.(3)Develop a Poisson-based travel demand distribution prediction framework.Considering the characteristics of traveler arrivals,we assume that the travel demand in each region and each time interval follows a Poisson distribution,centered on real demand.Then,we develop a neural network with a tailored softmax layer to approximate this Poisson distribution and real demand.A multi-objective function is designed to simultaneously achieve two goals:minimize the KullbackLeibler(KL)divergence between the distribution generated based on the real demand and the distribution predicted by the neural network;minimize the difference between true demand and the expectation of the predicted distribution.Based on experimental evaluations on taxi order data from Shenzhen and New York City,the proposed distribution forecasting framework achieves higher demand range forecasting accuracy,outperforming the baseline model in terms of reliability.Figures:31,tables:14,references:94...