Research On Traffic Condition Feature Mining Based On Multi-source Information Fusion And Spatiotemporal Modelin

In the past three decades of urban development,Intelligent Transportation Systems(ITS)has fully applied modern computer technology and communication technology in the field of transportation.It has greatly promoted and optimized the development of modern transportation modes and urban road networks.With the realization of multi-source and large-scale intelligent transportation systems,complex traffic information interaction and data expansion have laid a solid foundation for the research of multi-source traffic big data.Similar to the traditional model of big data research,multi-source traffic big data also exists in the research of structured data,semi-structured data,and mixed data.Its research also involves the collection,management,processing,and visualization of data.The research also involves the processes of data collection,management,processing,and visualization.This paper takes multi-source information fusion and a deep spatial-temporal model in the field of intelligent transportation systems as the basic research direction.Besides,this paper discusses the advantages and disadvantages of existing research solutions.Then,this paper proposes different data design models and model design algorithms for the combination of traffic multi-source information fusion and three application cases(non-regular structure,graph structure,and tree structure).Finally,the Tree Convolutional Neural Network is indicatively proposed to solve the data relevance imbalance problem caused by small-scale aggregated nodes.The data standardization and fusion process are carried out in this paper for complex and variable traffic multi-source information.According to the characteristics of different traffic applications,this paper divides the application cases into non-regular structures and regular structures.The non-regular structure mainly involves the grid division of urban traffic areas.Besides,the regular structure mainly involves the graph structure design and tree structure design of urban traffic roads.In order to deepen and mine the spatial-temporal variation features of the data,this paper designs a spatial correlation capture module for spatial distribution relationships and a temporal-dependent capture module for temporal data changes.This paper combines various deep learning model structures and data normalization methods to realize the task of predicting traffic flow,traffic speed,and road occupancy.They represent the characteristics of traffic conditions.The innovations and contributions of this paper are as follows.For the traffic condition mining task in non-regular structured application cases,this paper uses the design pattern of urban grids to assign regularized features to traffic data.Besides,this paper proposes a deep spatial-temporal prediction model.The model is designed with a spatial-temporal convolution module for neighboring grid cells,which uses a special convolutional neural network to capture the spatial correlation features and temporal-dependent features between grid cells.Compared with the current studies,the average improvement probability of this work is 11.14% on MAE and 11.91% on RMSE.For the task of traffic condition mining in the rule structure application cases,this paper designs a deep graph spatial-temporal prediction model using the spatial distribution information of traffic road data monitoring nodes.The model utilizes not only the ability to capture the spatial location relationship of nodes and their edges using the convolution calculation of the graph,but also enhances the model’s ability to perceive and fuse multi-source information by using the COVID-19 epidemic data as multi-source data.Compared with the current studies,the average improvement probability of this work is 51.01% on RMSE,53.63% on MAE and 56.34% on MAPE.For the traffic condition mining task in the rule structure application cases,this paper designs a Deep Tree Traffic Condition Prediction model for small-scale aggregation node distribution.To enhance the spatial correlation features of the tree structure nodes,the model proposes a Spatial Tree Convolution module and increases the contextual information features of the temporal data by using a cascade structure of multi-layer residual units.Compared with the current studies,the average improvement probability of this work is 24.35% on MAE and 40.96% on RMSE.