基于聚类算法和图神经网络的短时交通流预测

Short-term traffic flow prediction based on clustering algorithm and graph neural network

针对现有预测模型未充分利用交通流数据时空相关性的问题,本文提出了一种深度学习模型,将聚类算法、图神经网络(GNN)和门控循环单元(GRU)相结合。首先,利用聚类算法将预处理后的数据划分为不同类型的流量模式;其次,采用GNN提取复杂路网交通流的空间相关性,结合道路的皮尔逊相关性分析和节点的局部聚类系数,挖掘潜在的节点连接关系;再次,使用GRU提取交通流数据之间的时间相关性,通过自注意力机制捕获数据之间的相互依赖关系;最后,通过残差连接将GRU和GNN的输出与原始输入结合,经过全连接层得出最终的预测结果。多组实验结果证明,本文提出的模型预测精度优于其他基线模型及对比的模型。

Aiming at the problem that existing prediction models fails to fully utilize the spatio-temporal correlation of traffic flow data,this paper proposes a deep learning model that combines clustering algorithm,graph neural network(GNN)and gated recurrent unit(GRU).First,the algorithm classifies to classifies preprocessed data into traffic patterns;then,the GNN is used to extract the spatial correlation of the traffic flow of the complex road network,integrating Pearson correlation analysis of the roads and the local clustering coefficients of the nodes to uncover potential node connections;the GRU is used to extract the temporal correlation between the traffic flow data,and through the mechanism of self-attention,captures interdependencies among data;finally,the outputs of GRU and GNN are combined with original inputs via residual connectivity,and the final prediction results are obtained after the fully connected layer.Multiple sets of experimental results demonstrate the superior prediction accuracy of the proposed model is over other baseline models and contrast model.