Research On Deep Learning Prediction Of Dynamic Online Car-hailing Demand Considering Spatio-temporal Sequences

Online car-hailing demand prediction can provide decision support for rational and efficient scheduling of online car resources in the construction of smart cities,which is of great practical significance for balancing the supply and demand of urban online car-hailing resources in the spatial and temporal dimensions,relieving traffic congestion,and improving the utilization rate of online car-hailing resources.Most relevant studies have only attempted to capture the complex spatio-temporal patterns of demand for online vehicles,but ignore the intrinsic influence of regional functionality and failing to effectively model the dynamic temporal periodicity of demand for online vehicles.In this paper,after fully investigating and analyzing the shortcomings of the previous studies related to the demand forecasting of online vehicles,a highly accurate and efficient solution,i.e.,BDSTN(BERT-based Deep Spatial-Temporal Network)model,is proposed in a targeted manner,which enables high precision and fine-grained prediction of the demand for online car hailing in the city.Thereafter,this paper extends BDSTN and proposes a multi-task temporal prediction model paradigm in cities,i.e.,MT-BDSTN.The main studies of this paper are:(1)The paper divides the historical demand series of online car-hailing into long-term,medium-term,and near-term phases to explicitly model its dynamic time periodicity,and models the functional similarity among regions using POIs(Points of Interest).(2)A BERT-based deep spatio-temporal network model BDSTN is proposed,and experimental results on real-world dataset demonstrates that the proposed model outperforms other related algorithms in terms of RMSE and MAPE,and also verifies that its average prediction speed is at least 58.97%higher than other deep-learning-based algorithms on processors that supporting parallel computing.(3)Based on BDSTN,a multi-task learning paradigm is proposed based on PLE for finegrained temporal forecasting tasks in cities,and experiments are conducted on real-world dataset with two forecasting tasks,i.e.,the picking-up demand for online cars and the demand for drop-offs,to demonstrate that the paradigm can save computational and storage costs without significantly affecting the forecasting effect of each target,and is highly flexible and scalable.