Research On Urban Land Function Identification And Traffic Flow Prediction Using Spatio-temporal Graph Neural Networks

Having speeded up urbanization process,the city layout changes rapidly while its size expands dramatically.Accordingly,the increasing information makes urban management challenging,the information held by managers also needs to be updated in time to ensure the quality of residents’ life.Therefore,it is extremely important to design machine learning algorithms to assist managers in urban planning via exploiting big data generated from the daily life.Among current studies of assisting urban managers in planning,there are two branches,i.e.,the automatic identification of land function and the prediction of regional traffic flow.There is a close and complex correlation between land function and traffic flow,and such correlation information can be beneficial for the identification and prediction: on one hand,the traffic data generated by residents reveal potential information of land functions,which can be used to identify land functions;on the other hand,land function determines traffic demand affecting the traffic flow(e.g.,two regions that have similar variation patterns of traffic flow have the similar function).Therefore,we in this thesis will explore the correlation between above two aspects and design a visualization system for assisting urban planning.More details are listed below,(1)In view of the fact that the existing urban land function identification methods are not suitable for the development trend of complex land use functions in some areas,the similarity of regional functions cannot be further identified,this thesis designs and implements a deep learning model,namely RDYGE(Region Dynamic Graph Embedding)based on time series dynamic graph embedding.Based on the integration of Didi Chuxing order and Points of Interest(POI)data,the underlying spatiotemporal features of urban areas can be extracted,and combined with cluster analysis,the semantic identification of urban land functions can be realized.The experimental results show that the agglomeration degree of the same type of functional areas obtained by this method is high,and it can capture the changes in land use functions in complex areas under different time patterns better.(2)Aiming at the problem that the existing short-term traffic flow prediction methods in urban areas do not fully consider the causal relationship between regions,resulting in poor prediction accuracy,this thesis proposes a short-term traffic flow prediction algorithm based on location priors(Spatio-temporal Networks Based on location priors,ST-BLP).First,the algorithm uses a time series causal mining algorithm to build a traffic sequence causal graph.Then,combined with Graph Convolutional neural Network(GCN)and Gated Recurrent Unit(GRU)neural network,the spatial and temporal features between regions are extracted to predict the traffic conditions of the region in the next 5 to 15 minutes.Experiments on the Chengdu Didi Chuxing dataset show that this method outperforms other comparison algorithms in the prediction algorithm indicators,and can provide causal relationship analysis between regions.(3)Design and implementation of a visual analysis system for urban development based on multi-source data.The system integrates the urban land function identification algorithm and short-term traffic flow prediction algorithm proposed in this thesis,expresses the results of the algorithm intuitively,and provides users with an interactive operation interface.In conclusion,the thesis studies the two tasks of automatic identification of urban land function and accurate prediction of regional traffic flow in turn.The proposed method has achieved certain improvements compared with other comparative methods in the experiment of Chengdu as an example.Finally,the visualization system is designed and developed by combining the two methods,which has important practical significance for assisting the work of urban planners.