基于交通因子状态网络的城市交叉口交通流预测

信息技术的快速发展，为交通研究和城市交通管理提供了大规模、多样化的数据资源，并为城市交通状态估计和交通流预测方法的研究提供了有力支持。将城市交叉口视为一个微观交通系统，采用数据驱动与领域知识结合的方式，建立微观层次的交通因子状态网络模型（Traffic Factor State Network，TFSN），考察交通因素之间的相互关联，并考虑环境因素的影响。该模型结合交通因子和环境影响因子的影响，通过对交通流数据进行聚类分析，估算出对应于环境影响因子的交通状态，并通过实际案例验证其物理意义以及与交通流实际状态的对应关系。进一步地，基于不同交通状态下的交通流数据建立高阶多元马尔可夫链，进行交通流预测，并根据交通流时间序列的聚类性能指标提高模型的预测准确性。对数据序列马氏性强弱、马尔可夫模型阶数与模型预测准确性之间关系进行分析。研究结果表明：根据马氏性合理选择马尔可夫模型的阶数可以提升模型预测准确性；直接对原始交通流数据进行预测的平均绝对百分比误差为24.61%，而不同交通状态下交通流预测的平均绝对百分比误差为16.99%，相比直接预测误差下降了7.62%，验证了所提出的微观交通因子状态网络的有效性和可用性。

Traffic Flow Prediction of Urban Intersections Based on a Traffic Factor State Network

The rapid development of information technology provides diversified and large-scale traffic data resources for traffic research and urban traffic management, as well as strong support for research on urban intersection traffic state estimation and traffic flow prediction methods. In this study, urban intersections were regarded as a microscopic transportation system, and a combination of data-driven methods and domain knowledge was used to establish a microscopic traffic factor state network (TFSN) model to examine the correlation between traffic factors and consider environmental factors. Combined with the influence of traffic factors and environmental impact factors, this model estimated the traffic state corresponding to environmental impact factors through cluster analysis of traffic flow data and verified the physical significance and corresponding relationship of the traffic state with the actual traffic flow state through a real-world case. Furthermore, based on the traffic flow data of different traffic states, a high-order multivariate Markov chain was established to predict the traffic flow, and the prediction accuracy of the model was improved by clustering the performance index of the traffic flow time series. By analyzing the strength of the Markov property of the data, and the relationship between the order and the prediction accuracy of the Markov model, it was concluded that the prediction accuracy of the Markov model can be improved by reasonably choosing the order of the Markov model. The results show that the average absolute percentage error of the direct prediction of the original traffic flow data is 24.61%, while the average absolute percentage error of the traffic flow prediction under different traffic states is 16.99%, which represents a reduction of 7.62%. This verifies the effectiveness and availability of the proposed microscopic traffic factor state network.