Research On Traffic Flow Prediction Method Based On Graph Convolutional Neural Network

With the development of society,the number of cars continues to increase,and traffic congestion has gradually become the main problem hindering urban development.Traffic flow data records the number of traffic entities passing through a certain section of the road for a period of time,and its real-time and accurate prediction plays an important role in improving traffic congestion.However,due to the strong dynamics and complexity of the traffic system,it is difficult to get a relatively accurate forecast value.Therefore,it is of great significance to study the theory and method of traffic flow forecasting to better utilize the performance of urban traffic road network.Nowadays,recurrent neural networks(RNNs)quickly occupy an important proportion in time series forecasting methods by virtue of its memory function,and it is also one of the basic methods to capture the temporal characteristics of traffic flow data.After the emergence of graph convolutional network(GCN),it has been widely used to capture the spatial characteristics of traffic flow data by virtue of its advantages in processing topology.The combination of RNNs and GCN has therefore become a common method for traffic flow prediction at the comprehensive spatio-temporal level.However,the demand for long-term prediction and the lack of adjacency relationships have brought new challenges to the task of traffic flow prediction,existing models cannot dig deep into traffic flow.The complex spatial correlation and dynamic temporal correlation between stream time series often fail to achieve the required prediction accuracy.This thesis comprehensively uses deep learning methods such as graph convolution,one-dimensional convolution,autoregressive,and attention mechanism to build a traffic flow prediction model and conduct experimental verification.The experimental results show that the final traffic flow prediction model has achieved better than the benchmark the predictive performance of the model.The main research work of this thesis are as follows:1)Aiming at the problem that traditional traffic flow prediction methods cannot deeply capture the spatiotemporal characteristics of traffic data,a traffic flow prediction model based on graph convolution and sequence to sequence(Seq2seq)is proposed.First,the correlation between traffic flow sequences is calculated by graph convolution based on temporal similarity to model spatial dependencies,and then a sequence-tosequence model composed of gated recurrent units(GRUs)is used to model temporal dependencies.And for the case that there are more error data in the data set,the improved loss function is used for model training to improve the over-fitting problem of the model and increase the robustness of the model.The data set is selected for experiments to verify that the model has an improved effect on traffic flow prediction.2)Aiming at the problem that the recurrent neural network cannot be calculated in parallel and the traditional graph convolution cannot make full use of the spatial information of traffic data,a traffic flow prediction model based on k-shape clustering and spatio-temporal convolutional neural network is proposed.First,the k-shape clustering algorithm is used to cluster the traffic flow time series to calculate the semantic adjacency relationship,and then the semantic adjacency relationship and the spatial adjacency relationship are fused to obtain a composite adjacency relationship,and the final adjacency matrix is obtained by applying the attention mechanism.After the adjacency matrix is obtained,it is input together with the traffic data into the stacked spatio-temporal convolution module,and the spatio-temporal characteristics of the traffic data are captured through gated temporal convolution and graph convolution,and then the final prediction result is obtained.The data set is selected for experiments,and compared with known traffic flow prediction methods,it is verified that the model has an improved effect on traffic flow prediction.3)The fixed adjacency matrix cannot capture the dynamic correlation between traffic sequences,and the self-attention mechanism shows outstanding advantages in perceiving long-term dependencies.Based on this,a traffic flow prediction model based on convolutional self-attention mechanism and dynamic graph convolution is proposed.Firstly,the dynamic laplacian matrix generated based on node similarity is used in graph convolution to deeply mine the relationship between nodes.The spatial correlation between them;then the local temporal correlation is captured by the convolutional neural network,and the long-term correlation is captured by the selfattention mechanism.In addition,in order to improve the over-smoothing problem of graph fusion and deep network,an autoregressive module is introduced through a special combination method,which enhances the model’s perception of aperiodic changes in input data.That is to say,in traffic flow forecasting,this model not only considers the dynamic relationship between multiple time series,but also considers the long-term,short-term,periodic and non-periodic changes of a single time series.For this module,four classic data sets are selected for simulation experiments,and compared with known traffic flow prediction methods,it is verified that the model has an improved effect on traffic flow prediction.