Research On Prediction Method Of Ride-Hailing Demand Based On Spatio-Temporal Graph Neural Network

With the increasing usage of mobile Internet technology,the ride-hailing market has rapidly expanded in recent years.Nowadays,ride-hailing has become one of the most popular modes of people’s daily travel.The healthy development of ride-hailing can serve diversified travel needs and increase idle vehicle utilization.Furthermore,it plays a prominent role in providing people with flexible employment options as well as relieving many problems in the urban traffic environment.However,there are many bottlenecks on the ride-hailing,such as the long taxi queue of passengers during peak hours,the long no-load time of drivers during normal hours,and the imbalance between the supply and demand of passengers and vehicle resources.To overcome the above limitations,we conduct research on ride-hailing demand forecasting based on spatio-temporal graph networks in a data-driven manner.The main research contents are listed as follows:(1)A multi-scale spatio-temporal convolutional network model based on the attention fusion mechanism is proposed to predict the ride-hailing demand in a single area.Firstly,multi-scale time series sequences are constructed to describe the periodic relationship in the historical demand sequences.With the residual spatio-temporal convolution module,the temporal and local spatial features of the sequences on different times scales are extracted quickly and effectively.Secondly,in terms of feature fusion,a spatiotemporal feature fusion module based on the self-attention mechanism is proposed.The module considers the influence of external factors such as weather and holidays on the prediction results and fuses the extracted spatio-temporal features with the self-attention mechanism.Finally,the graph convolution module performs graph convolution operations on the constructed POI similarity graph to capture global contextual semantic dependencies.The experimental results on two real datasets show that proposed model outperforms the other nine baseline models.(2)A multi-graph convolutional network model based on the internal temporal attention mechanism is proposed to predict the demand flow of ride-hailing in different areas.Firstly,an internal temporal attention mechanism module is designed.This module captures the effects of different time scales sequences and external factors on the prediction performance by calculating the attention weight.Secondly,in terms of capturing spatial dependencies,a multi graph convolutional module is designed.This module builds three graph as spatial neighborhood graph based on geographic neighborhoods,contextual semantic graph based on similarity of interest points,and dynamic sequence graph based on historical sequences.Moreover,graph convolution operations on the constructed graphs are applied to capture the spatial dependencies from both static and dynamic perspectives,respectively.The experimental results on two real datasets demonstrate that proposed model outperforms the other seven baseline models.