基于图网络融合的交通状态预测方法研究

Research on Traffic State Prediction Method Based on Graph Network Fusion

文中考虑道路节点之间的时间相关性,利用皮尔逊相关性系数构建逻辑相关路网;通过图聚合算法聚合道路节点邻居信息,融合原始交通路网与逻辑相关路网提取的时空特征信息,以最小化损失函数为目标,返回最优模型参数,构建基于图网络融合的交通路网模型.采用西雅图高速路网速度数据集(seattle)和加州流量数据集(PEMS08)作试验验证,图网络融合模型提高了在交通状态预测精度.在短时交通状态预测中,Seattle的MAE指标为2.57、MAPE指标为6.48;PEMS08的MAE指标为14.23、MAPE指标为7.15;长时交通状态预测结果均优于LSTM、T-GCN等模型.

Based on the time correlation between road nodes,Pearson correlation coefficient was used to construct a logically related road network.The neighbor information of road nodes was aggregated by graph aggregation algorithm,and the spatio-temporal feature information extracted from the original traffic network and logically related road network was fused.With the goal of minimizing the loss function,the optimal model parameters were returned,and the traffic network model based on graph network fusion was constructed.Seattle expressway network speed data set(Seattle)and California traffic data set(PEMS08)were used for experimental verification,and the graph-network fusion model improved the accuracy of traffic state prediction.In the short-term traffic state prediction,Seattle’s MAE index is 2.57 and MAPE index is 6.48.The MAE index and MAPE index of PEMS08 are 14.23 and 7.15 respectively.Long-term traffic state prediction results are better than LSTM,T-GCN and other models.