Research On Traffic Flow Prediction And Congestion Detection For Freeway Through Spatio-temporal Toll Data Mining

The toll collection scenario on freeways is a large-scale complex nonlinear system characterized by a wide domain range and significant fluctuations in traffic flow.Accurately predicting the spatiotemporal traffic conditions and preemptively detecting congestion in the area holds important theoretical and practical value for enhancing freeway operational efficiency and future intelligent traffic management.With the significant advancements in perception,communication,and fusion modeling technologies,extracting critical traffic information and spatiotemporal relationships from massive high-dimensional heterogeneous data has become a key challenge in the digital development of freeways.This thesis focuses on toll collection gates and ETC gantries as research objects,employing advanced modeling approaches such as deep learning and fusion theory to address key issues in application scenarios.The research unfolds along the logical line of "background condensation,data analysis,feature extraction,prediction design,and congestion identification." The main contributions and achievements of this paper can be summarized as follows.Firstly,from various perspectives such as data investigation,quality issues,and anomaly repair,the research highlights toll collection gates and ETC gantries as two research scenarios,delving into the quality issues in toll data.It designs an anomaly detection approach for different quality issues and proposes a multi-class generative adversarial network(MC-GAN)algorithm for missing data imputation.The MC-GAN model achieves optimal completion performance under different missing ratios,particularly exhibiting an improvement of approximately 14.67%compared to the previous approach in high missing ratio scenarios.This provides a systematic analytical approach and repair method for abnormal traffic data.Secondly,utilizing correlation analysis and data visualization methods,the research explores the traffic flow characteristics and regional prediction modeling approach in the toll collection scenario from a spatiotemporal dimension.It introduces a multi-graph coding spatiotemporal dilated convolution network(MGC-STDCN)based on a deep learning framework.The model employs graph theory to restructure the large-scale road network topology into multiple types of non-Euclidean graph structures,representing the real spatial dependency between different observation sites.It overcomes the limitations of existing methods in capturing long-term sequences in large-scale and parallel manners.Experimental designs in five sections validate the proposed model’s ability to achieve good long-term prediction performance and computational efficiency in a large-scale traffic network context.Lastly,taking the toll collection gate as an example of a complex traffic scenario,the research designs congestion domains,selects traffic discriminant indicators,and redesigns indicator calculation methods based on the composition characteristics of toll collection gates.It proposes an improved fuzzy clustering algorithm by incorporating entropy weight and sample weight concepts.Using this approach,it classifies toll collection gates into congestion levels and then combines it with the MGC-STDCN prediction model to achieve accurate identification of congestion states in real toll scenarios at future time points,providing decision support for freeway traffic congestion management and road network planning.