Taxi Passenger Demands Prediction Based On Deep Learning Approach

Passenger demands prediction is an important part of the intelligent transportation system which cannot be ignored.An accurate predictive model can help the online car-hailing system to distribute orders and dispatch vehicles more rationally,thus solving the contradiction between supply and demand.This will not only reduce energy waste but also effectively alleviate urban traffic congestion.The popularity of online car-hailing systems like DiDi and Uber has enabled us to consistently collect large-scale passenger demand data,and how to use these big data to improve the accuracy of demand prediction is an interesting and critical issue.Traditional demand prediction methods are mostly based on time series prediction techniques,which cannot effectively model nonlinear spatio-temporal relationships.Recently,deep learning technology has made breakthroughs in the fields of computer vision and natural language processing,which greatly inspired us to use deep learning technology to solve traffic prediction problems.Most of the existing traffic prediction methods only consider spatial correlation or time correlation independently,and are mainly for single-step prediction scenarios.In this paper,the city space is viewed from a new graph perspective,and the spatial correlation and time correlation are modeled simultaneously using structured recurrent unit.A hybrid spatialtemprol graph convolutional recurrent network ST-HGCRN based on Encoder-Forecaster architecture is proposed.It can simultaneously predict the demand of passengers getting on and off from n irregular areas of the city at the next k moments.Our work can be summarised as following three points:(1)A spatial correlation modeling method based on high correlation nodes is proposed.We first quantify the demand correlation between urban areas from different perspectives,with multiple regional relationship graphs generated;And we then only retain those nodes with higher correlation in these graphs,and finally use the hybrid graph convolution to integrate the information of high correlation nodes.Our approach makes more efficient use of the demand interaction between locations than CNN-based approaches.(2)Combined with the structured recurrent unit HGCRU and Encoder-Forecaster architecture,We have solved the multi-step demand prediction problem.We redesigned recurrent unit of the RNN using hybrid graph convolution so that the RNN can take advantage of the spatial structure of the data,enabling simultaneous modeling of spatial correlation and temporal correlation.HGCRU has two specific instance: HGCLSTM and HGCGRU.With these two HGCRU instances and the Encoder-Forecaster architecture,we proposed the STHGCRN network,which can solve the multi-step demand prediction problem effectively.(3)We conducted a number of experiments on a large-scale order dataset of Chengdu provided by DiDi.The performance of models as well as the influence of various parameters on them were both analyzed.Our model won the comparison model in most cases,achieving more than 10% performance improvement and showing considerable competitiveness.Various self-contrast experiments also demonstrate the importance of the various components in the network.