Research On Urban Crowd Flow Transfer Learning Based On Regional Spatio-temporal Similarity

Crowd flow in a region is a description of the flow of groups in a certain space.Prediction of crowd flow can be used to analyze urban traffic conditions,to relieve traffic congestion,to maintain public safety,to design urban layout,etc.It can also help people understand the functional areas of the city and people’s activity patterns,etc.It has great application value in traffic operation and maintenance,disaster response,tourism recommendation,urban planning,etc.Deep learning has been successfully applied to solve this problem,but it relies on rich historical data.In fact,this prediction approach,relying on rich historical data,may suffer from data scarcity due to the current inconsistent modernization and digitization of cities,which face serious data scarcity for reasons such as the slow digital development process of cities or the government’s development plan resulting in the corresponding pedestrian flow data being restricted by privacy protection and the temporary unavailability of massive urban traffic data.To overcome this problem,this thesis has proposed a new transfer learning method for urban crowd flow based on regional spatio-temporal similarity.(1)A spatial information entropy model based on multi-source data is proposed,and a similar region matching algorithm is constructed to lay the research foundation for the transfer learning process to find the appropriate source and target regions.The spatial information entropy model of regional structure is formed by taking administrative divisions as regional units,considering the point,line and surface elements of urban spatial structure,combining the information entropy of POIs,road network data and building data,and then assigning weights to them based on the AHPentropy weighting method to quantify the similarity of urban structure between regions.At last,6districts in the core of Beijing are used as the research area to compare the regional similarity between them,and the results of the experiment are consistent with the reality of Beijing,verifying the accuracy and reliability of the urban structure similarity method.(2)A deep spatio-temporal neural network structure,named Region Rep,is constructed based on Conv LSTM(convolutional long and short-term memory neural network)for similar region pairs,using the external environmental data of the region as the regional features,designing a hidden layer specifically for the representation of regional features,and using Region Rep to train on its rich historical data to obtain the crowd flow prediction of the source region.At last,using Dongcheng District as the source region,using its rich historical data of bike-sharing to represent the crowd flow,after the training of the model and the adjustment of parameters,a high-performance crowd flow prediction model of the source region is obtained.(3)A cross-regional transfer learning algorithm is proposed to further optimize the crowd flow prediction model to achieve crowd flow prediction in the target region where data is scarce,completing the transfer learning study.A region-based cross-city transfer learning algorithm is designed to migrate and apply the source region prediction model to the target region by minimizing the differences in external environment feature representation between region pairs.Finally,the experiments of crowd flow prediction in Dongcheng District(source region)and Xicheng District(target region)are completed,and after verification and analysis,the crowd flow prediction in the target region can be achieved with high accuracy.