Online Car-hailing Demand Forecast Based On Graph Convolution And Fully Connected Model Fusion

With the rapid development of the Internet,online taxi service platforms such as Didi Dache came into being.Online car-hailing is becoming more and more popular with people because of its convenience and flexibility.Online car-hailing is also playing an increasingly important role in urban transportation.Establishing an online car-hailing demand forcast model can proactively optimize the disptatch of online car-hailing resources,increase the utilization rate of online car-hailing,reduce the waiting time of users,and at the same ti me alleviate urban congestion to a certain extent.This thesis focuses on the forecast of online car-hailing demand.The main research content includes:1)In the data preprocessing stage,the administrative district data is applied to divide the city into irregular areas,thus avoiding the problems of low practical value or difficulty in the existing spatial partition methods.On this basis,the irregular area is regarded as the node in graph theory,the demand for car-hailing in the region is regarded as the feature of the node,and the dependence of the demand for car-hailing between regions is regarded as the edge with weight between the nodes.The problem of online car-hailing demand prediction is transformed into a graph regression problem.2)In the feature selection stage,Aiming at the problems of insufficient feature extraction and insimplification of features in existing research papers,features are extracted from two perspectives,space-time and globality.Three time key step features are extracted in time;three spatial dependency matrixes are constructed in space;global features such as weather conditions are extracted at the same time,and treature selection is performed based on the random forest model.3)In the feature modeling and prediction stage,a multi-graph convolution unit is first proposed to improve the input structure of the graph convolution network.and then,based on the improved graph convolutional network and the fully connected network,the spatial-temporal and global characteristics of online ride-hailing vehicles were modeled respc,crcively.Finally,the graph convolutional network and the fully connected model were fused to establish the demand prediction model of online car-hailing,which improved the problem of low prediction precision of the existing single model.This thesis uses the Keras deep learning framework to analyze and verify on real data sets.Comparing the model proposed in this thesis with the current mainstream model,the root mean square error and average absolute error indicators are 24.9%and 13.0%respectively lower than the current best model,showing that the proposed model has certain advantages.At the same time,the functions of the three key time steps,three spatial dependence features and the global features of the proposed model are verified.