Research On Traffic Flow Forecasting Method Based On Deep Spatial-temporal Models

Since the reform and opening up,people’s living standards have been greatly improved,and the demand for motor vehicles has increased year by year.Although motor vehicles can facilitate people’s travel,the rapid growth of their numbers has increased traffic congestion on urban roads.Traffic flow prediction is an indispensable part of intelligent transportation system,which can not only help travelers make better travel decisions,but also assist traffic management departments to formulate more effective management methods.Therefore,accurate and effective traffic flow prediction is an important basis for improving the quality of traffic management services.Traffic flow prediction is to predict by mining the temporal or spatial characteristics of historical traffic flow data,and its research methods are mainly based on statistical methods,machine learning methods and deep learning methods.However,most of the existing methods cannot comprehensively and accurately capture the spatiotemporal dynamic correlation of traffic flow in the road network,resulting in low traffic flow prediction accuracy.Therefore,this thesis uses deep learning methods to fully learn the spatiotemporal dynamic correlation characteristics of traffic to improve the accuracy of traffic flow prediction.The main research work and contributions of this paper are as follows:1.Aiming at the problem that most of the existing models capture the temporal dynamic correlation of long-term traffic flow series and are prone to gradient disappearance or gradient explosion,this thesis proposed a traffic flow prediction model STDCNN based on convolutional neural network.The model introduces a dilated causal convolutional neural network,which can capture the temporal dynamic correlation of traffic flow with fewer network layers and learning parameters,and then uses a graph convolutional neural network to extract the local spatial correlation of traffic flow in a road network with a topology structure.Finally,the final prediction result is obtained through the fully connected layer.The experimental results show that the model in this thesis can not only obtain better prediction accuracy,but also the introduced dilated causal convolution can speed up the model training.2.Aiming at the problem of incomplete local spatial dynamic correlation characteristics of traffic flow in existing models,based on STDCNN model,this thesis proposed a traffic flow prediction model STACNN based on spatial attention mechanism.The model introduces a spatial attention mechanism to dynamically focus on the relationship between non-adjacent traffic sensor nodes,and combines graph convolutional neural networks to capture the local spatial dynamic correlation of traffic flow.Then,gradient descent is performed using the charbonnier loss function of laplacian pyramid super-resolution network model to make up for the lack of the L2 loss function.The experimental results show that both the introduced spatial attention mechanism and the charbonnier loss function have a positive impact on the model,which further improves the prediction accuracy of traffic flow.