Research On Delivery Time Prediction Based On Spatio-temporal Trajectories

In recent years,the logistics industry has been expanding rapidly.The number of couriers has exceeded 3 million,while the daily delivery packages have exceeded 100 million in China,which brings great challenges to the logistics platform.Delivery time prediction(i.e.,to predict the arrival time of undelivered packages at any time)is one of the most important tasks in the logistics platform.On the one hand,predicting the delivery time is beneficial to the staff scheduling for company and route planning for couriers,thus the delivery efficiency can be improved.On the other hand,accurate prediction of delivery time can provide customers more punctual service and alleviate customers' waiting anxiety.Benefited by the widely adoption of electric devices,massive spatio-temporal trajectories of couriers in the delivery process is continuously accumulated,providing the data foundation for researches in the logistics field.The goal of this paper is to accurately predict the delivery time of the packages by learning from historical spatio-temporal trajectories.In real application scenarios,delivery time prediction is affected by a variety of complex factors including the courier and the delivery environment.And the status of the package delivery changes over time.In addition,unlike traditional one-destination arrival time prediction problem,arrival time of all the undelivered packages of a courier should be predicted simultaneously,which is essentially a multi-destination prediction problem.The above reasons bring huge challenges to the delivery time prediction.To deal with the above challenges,this paper proposes a delivery time prediction model based on spatio-temporal trajectories,STDTN,to learn to estimate the arrival time of package delivery from couriers' massive historical spatio-temporal trajectories.STDTN can learn the behavior patterns of couriers and capture the temporal and spatial dynamics in the delivery process.A spatio-temporal component with geographic information encoder,convolution operation and Bi-LSTM is proposed to capture the spatio-temporal information in the trajectories and model the dynamics in package delivery status.Moreover,a location-based selection strategy that is used to select a combination of historical similar packages obtains the representation of the undelivered packages,which integrates multiple characteristic information to complete the delivery time prediction.It is found that the delivery sequence has a huge impact on the multi-destination delivery time prediction.On the basis of STDTN,a Multi-Task model for Delivery Time(MTDTN)prediction is proposed.In addition to modeling the complex factors that can affect the delivery time and the delivery route,this paper uses the multi-task learning to simultaneously predict the delivery time as well as the delivery sequence.Based on the attention mechanism,the feature extraction of the collection of similar undelivered packages is performed,and fusion complex factors are fed into full connected layers to get the delivery sequence.The model performance is thereby enhanced by introducing the delivery sequence prediction as an auxiliary task.Finally,extensive experiments are conducted on a real-world delivery dataset.The experimental results show that the STDTN model proposed in this paper performs better than the existing methods,and the MTDTN model has further improvement in delivery time prediction performance.