Research On Traffic Forecasting Method Considering Complex Spatial-temporal Dependence And Its Analysis Based On Visualization

With the progress of urbanization and the improvement of city dweller’s living standards,the number of automobiles has increased sharply.However,the existing transportation infrastructure cannot meet the rapid growth of the number of vehicles,which leads to the increasingly serious traffic congestion problem.With the development of sensing,communication and computing technology in the field of intelligent transportation,it is a challenging task to mine the potential traffic state change patterns and characteristics from the massive,high-dimensional and diversified traffic spatio-temporal data,so as to predict the future traffic state and provide efficient and accurate information service for traffic management departments as well as the public,and to mitigate the traffic congestion issue ultimately.This paper analyzes the status quo and relevant knowledge of traffic flow prediction.Given the complex characteristics of traffic flow data and the shortcomings of existing forecasting methods,this paper proposes a deep-learning-based traffic prediction method from the perspective of data-driven models.This model considers both complex spatial-temporal dependence and traffic flow theory.On top of that,a dynamic visualization system of spatial-temporal dependence is implemented driven by the need to interpret the temporal and spatial dependence learned by the prediction model.The specific work and contribution are as follows:Firstly,a multi-step traffic flow prediction model considering complex spatial-temporal dependence is proposed,namely STSeq2 Seq.In terms of spatial dependence modeling,the pattern-aware adjacency matrix(PAM)is proposed,and the graph convolution neural network is used to extract the dynamic nonlocal spatial correlation of traffic flow;in the aspect of temporal dependence modeling,the sequence to sequence model of ”convolutional encoder + recurrent decoder” is proposed to effectively capture the dynamic temporal dependence between historical and future traffic series.Finally,all the components are integrated to achieve the modeling of complex spatial-temporal dependence and accurate prediction of future traffic flow.Secondly,extensive experiments are conducted on two public datasets.By comparing the proposed method with other eight baseline methods,we prove the superiority of the proposed algorithm with respect to multi-step prediction performance.The effectiveness of the proposed algorithm in spatiotemporal dependence modeling is verified by hyper-parameter influence experiment and ablation study.Finally,through the comparison of the running times of different algorithms,it shows that the proposed algorithm has relatively high training and inference efficiency.Thirdly,this paper designs and implements the dynamic visualization system of spatiotemporal dependence for the proposed algorithm.The Web-based system is realized to fulfil the requirements for model interpretability and spatial-temporal dependence analysis.The functions of the system are tested and several case studies are carried out on the visualization results to verify the interpretability of STSeq2 Seq and lay a solid foundation for the practical application of the algorithm.