Research On Taxi Idle Time Prediction Based On Trajectory Data

In recent years,with the continuous expansion of urban scale,the number of taxis and online car-hailing has increased,and residents’ demand for online taxi-hailing has also increased.While the taxi market is rapidly increased,competition among industries is also becoming increasingly fierce,so how to better conduct route planning,reduce the empty-load rate and idle time of taxis is a key factor in improving driver’s profit.As an important measurement indicator,the taxi idle time reflects the utilization efficiency of taxi resources and the driver’s income status from one side.Accurate taxi idle time prediction can effectively guide drivers to make reasonable path planning,and assist taxi platforms for efficient resource scheduling.Taxi idle time prediction is a novel and complex problem.However,in actual scenarios,the idle time in different areas of the city is affected by various factors such as regional traffic,passenger flow,and historical idle time.Therefore,how to model this information,accurately predict the taxi idle time is one of the problems that the industry needs to solve.Therefore,this paper is based on the trajectory information and passenger records of the taxi,explores the spatial-temporal correlation of the taxi idle time data.By integrating various influencing factors and auxiliary tasks,it predicts the taxi idle time in various areas of the city,and provides route planning support information for the taxi platform.First,because the spatial-temporal relationship in taxi idle time data is complex and difficult to extract,a spatial-temporal connection graph convolutional network(STCGNN)model is proposed,which uses a novel spatial-temporal connection graph structure to model spatial-temporal data.It not only creates a spatial graph,but also connects multiple spatial graphs in the time dimension to make the graph structure.Performing the graph convolution operation based on the spatial-temporal connection graph can make the model simultaneously capture the spatial,temporal,and spatial-temporal relationships in the taxi idle time data.The temporal module is further designed,based on the structure of the recurrent neural network to capture the time sequence dependence that exists in the time series that the convolution structure is difficult to capture.Finally,the feature representation is input to output layer to get the final taxi idle time prediction result.Secondly,the STCGNN model still has shortcomings in graph convolution operations,external information utilization,and temporal components.Subsequently,several improvements were made to STCGNN,and Multi-Task Spatial-Temporal Similarity Graph Neural Networks(MSTSGNN)model was proposed.MSTSGNN adopts a novel spatial-temporal similarity graph structure.Based on spatial-temporal graphs,it improves the connection of time edges so that the graph structure itself contains more information.MSTSGNN further improves the graph convolution operation to make it suitable for spatial-temporal similarity graphs,which can capture the spatial dependence,temporal dependence,and spatial-temporal dependence in the data,and can distinguish the spatial-temporal heterogeneity in the spatial-temporal data.Replace the RNN of the temporal module with GRUGCN,so that it can keep the spatial structure information in the data while capturing the temporal information.A multi-task framework is introduced to learn the feature representation of data from different perspectives and make better use of taxi demand and taxi traffic data.An attention-based multi-task fusion mechanism is also proposed,which uses the attention mechanism to filter the information of auxiliary tasks to improve the information acquisition ability and prediction performance of the main task.Experiments were conducted on two Di Di public datasets in Chengdu and Xi’an.The experimental results show that the effect of the STCGNN model is significantly better than the existing prediction methods.Compared with STCGNN,the accuracy of MSTSGNN has been further improved.The experimental results show that the model proposed in this paper can effectively mine the temporal and spatial and temporal dependence of the taxi idle time data and accurately predict it.