Research On Traffic Flow Prediction Method Based On Convolution Neural Network

With the acceleration of China’s urbanization process and the continuous improvement of economic conditions,the number of motor vehicles has increased rapidly,the pressure on the urban road network has increased,and the problem of traffic congestion has become more serious.Intelligent transportation system effectively applies advanced information technology to transportation,thus establishing a real-time,accurate and efficient transportation management system to realize traffic control and guidance.Traffic flow prediction is an important component of intelligent transportation systems,and accurate and real-time prediction results are of great significance for optimizing traffic signal control,improving transportation efficiency,and enhancing traffic guidance capabilities in intelligent transportation.Traffic flow has strong nonlinear characteristics and spatiotemporal correlation,which makes traffic flow prediction very challenging.The existing traffic flow prediction models are difficult to effectively capture the spatiotemporal correlation features hidden in traffic flow data,and the modeling of spatiotemporal dependencies has certain limitations.In response to this issue,the main research content of this article is how to effectively capture the implicit spatiotemporal characteristics in traffic flow data and establish a reasonable prediction model.The main research of this article is as follows:(1)Traffic flow has complex random time-varying characteristics,which will have different impacts on future traffic states at different times and continuously change over time.A traffic flow prediction method TCN based on time convolutional networks is proposed for the random time-varying characteristics of traffic flow.TCN uses extended causal convolution to construct a temporal convolutional network layer to model the temporal characteristics of traffic flow,and uses gating mechanisms to control the information of the temporal convolutional network layer to capture the complex temporal dependencies of input data at different times.Increase the network depth by increasing the number of time convolution network layers,so that the Receptive field of convolution operation becomes larger,and further capture the deep time dependence of global time mode in data.The experimental results on two real datasets METR-LA and PEMS-BAY have shown that this method has a significant improvement in prediction accuracy compared to traditional statistical and machine learning methods.(2)How to obtain dynamically changing spatial relationships of road networks without prior knowledge is a challenge in current traffic flow prediction.A traffic flow prediction method ASTCNN based on attention and spatiotemporal convolutional neural networks is proposed to address the spatiotemporal dynamic changes of traffic flow.This method utilizes time convolution and attention mechanisms to fully explore the potential spatiotemporal dependencies in traffic flow data.By using a multi-layer time convolutional network,the long-term dependencies of traffic flow are captured.By using self-learning node embedding to construct spatial dynamic attention,it is possible to model the dynamic spatial correlation of different road network nodes without prior knowledge of the graph.A large number of experiments were conducted on four real traffic datasets(METR-LA,PEMS-BAY,PEMSD4,and PEMSD8),and ASTCNN achieved good prediction performance without input graph information.Moreover,in the multi-step prediction of 30 and 60 minutes,the prediction accuracy of ASTCNN is also very high.(3)A traffic flow prediction method STA-GNN based on attention and graph convolutional networks is proposed to address the complex global and local spatiotemporal dynamic changes of traffic flow.This method considers both spatiotemporal and global features.Firstly,the graph convolution network is used to capture the local spatial correlation in the road network topology structure,and a dynamic graph learning layer is proposed to capture the hidden one-way association between nodes by constructing a dynamic adjacency matrix.Then,the graph convolutional network aggregates node information with its neighbor information to extract node features.A multi-core time convolutional layer was designed to extract long-term and short-term local features,while capturing implicit multi-scale time features.Then,spatiotemporal attention is used to extract global temporal dependencies and spatial correlations,enhancing the representation ability of the model’s global spatiotemporal modeling.The experimental results on METR-LA and PEMS-BAY show that STA-GNN achieves good prediction performance in different datasets and prediction time periods.