Dynamic Spatio-Temporal Graph Convolutional Neural Networks For Online Car-Hailing Demand Forecasting

The rapid development of online car-hailing provides an excellent opportunity to meet the needs of convenient travel services.At the same time,as a new way of travel in smart cities,it has brought great convenience to urban residents’ travel,and drivers can also accurately pick up passengers.However,with the increase in the number of online car-hailing,urban road congestion will increase during peak traffic periods.Due to the randomness of passenger travel demand,it is difficult to balance the number of online car-hailing and passenger travel demand between supply and demand.In different environment,each region also has different online car-hailing travel needs,which are affected by multiple complex factors such as geography,regional functionality,meteorology,and travel habits.However,the accurate prediction of online car-hailing travel demand can effectively alleviate the imbalance between supply and demand,which has become a topic worthy of further study.Therefore,this work uses online car-hailing travel orders and other external data to analyze and study the spatio-temporal characteristics of online car-hailing travel demand.The model is constructed and verified for the travel demand prediction problem of online car-hailing,which improves the efficiency and accuracy of prediction.It is of great significance to improve passenger travel efficiency,optimize urban traffic resource allocation,and improve transportation capacity and road resource utilization.The main work and contributions are as follows:(1)The definition of the demand forecasting problem of online car-hailing is clarified.The Manhattan area is divided into regions by the method of grid division.Each grid represents a node.Based on the graph theory,the online car-hailing travel demand graph structure model is established.The dynamic spatio-temporal variation characteristics of online car-hailing travel demand are analyzed and modeled from the perspectives of temporal characteristics,spatial characteristics and other external influencing factors.The periodic spatio-temporal characteristics of online car-hailing travel demand are modeled,which provides a theoretical basis to improve the performance of online car-hailing travel demand prediction model.(2)A multimodal fusion graph convolutional network car-hailing travel demand prediction model(MFGCN)based on spatio-temporal dynamic graph convolutional neural network is constructed.The core idea is to employ the graph convolutional neural network structure to encode and learn the geographical,semantic and functional correlation information of the spatiotemporal characteristics of the online car-hailing travel demand between urban grids,which effectively solves the sparsity of the adjacency matrix.The multimodal attribute enhancement module learns the characteristics of weather factors and passenger travel habits,and improves the model’s ability to extract spatio-temporal nonlinear features.Finally,the periodic characteristics and time attention mechanism are combined to capture the fluctuation of different time sequences,and the periodic variation characteristics of online car-hailing travel demand are obtained,which further improves the prediction performance of online car-hailing travel demand.(3)Taking the real-world data of Manhattan in New York as a case study,the horizontal and vertical comparison experiments and prediction performance evaluation are carried out for the online car-hailing travel demand prediction model of the constructed multimodal fusion graph convolutional network.The results of the model are analyzed and evaluated from the perspectives of overall prediction results,adaptability of different prediction durations and anti-noise interference ability,which verifies the validity,adaptability and robustness of the construction method and model structure of the online car-hailing travel demand characteristics.The experimental results show that compared with other baseline models,the prediction accuracy of MFGCN model is improved by 6.74 %,and it has strong anti-interference ability.