Research On Traffic Flow Forecasting Based On Deep Learning

As a crucial component of intelligent transportation systems,accurate forecasting of traffic flow can provide more reliable macroscopic road condition information for transportation management departments and individuals,and provide important references for research on urban construction,road planning,traffic control and other issues.To fully explore the spatio-temporal characteristics of traffic flow data and address existing factors,this paper proposes three deep learning-based prediction models,aiming to further capture the elements in traffic flow data that contribute to the improvement of model accuracy and performance in terms of accuracy,stability,and other aspects.The main contributions of this paper are as follows:(1)A combination model of one-dimensional convolutional long short-term memory neural network based on self-attention mechanism is proposed.This model utilizes one-dimensional convolutional neural network to extract features from real road traffic flow.Then,it uses bidirectional long short-term memory neural network to model the time continuity and periodicity of traffic flow data from the perspective of time.Finally,the self-attention mechanism is used to redistribute the weight of the matrix.The model achieves the goal of enhancing the performance of data features and addresses issues such as weak feature perception or poor temporal modeling in the baseline time series prediction model.(2)Considering the spatial topology of the traffic network,this paper converts the nodes in the traffic network into a graph data structure and proposes a spatial-temporal graph convolutional long short-term memory neural network model based on spatial attention.The model first redistributes the weight values of the adjacency matrix representing the spatial connectivity between intersection nodes through the spatial attention mechanism.Then,it explores the characteristics of time and space dimensions in the spatio-temporal graph convolutional neural network.Finally,the time series is passed to the long short-term memory network for temporal modeling.This model enhances the prediction performance from the perspectives of spatial correlation and spatio-temporal dependence,and solves the problem that previous models can only process traffic flow data of a single traffic section in the planar dimension and cannot perceive the structure of the entire traffic network for changes in traffic flow at different intersection nodes.(3)A spatial-temporal graph convolutional long-short term memory network model is proposed that integrates an adaptive graph learning method to address the issue of the model’s inability to update connection relationships in real-time.This model can continuously update the optimal affinity matrix during the training process,dynamically adjust the model weight to adapt to the complex dependency relationships between nodes,and improve the overall performance of the prediction model.Finally,this paper conducted quantitative and qualitative experiments on the proposed three combination models using relevant data,and conducted parallel control experiments with baseline models.The results show that the three proposed combination models have considerable predictive performance for different data features,and have varying degrees of advantages compared to the prediction indicators of the baseline models.