基于机器学习的交通流预测方法综述

通过文献梳理、专家访谈和试验场景构建等方法，分析了道路指定断面和区域路网宏观交通流预测的国内外研究现状和发展趋势，归纳了局部断面交通流预测方法，包括传统机器学习、递归神经网络和混合模型，分析了卷积神经网络、图神经网络和融合多因素网络的特点，阐述了方法的原理、优势、局限性和应用场景，总结了现有场景交通数据集类别，从采样周期与采集方式角度归纳了国内外主流交通数据集。分析结果表明：递归神经网络可以有效获取交通数据的历史规律，但存在梯度爆炸、计算复杂度高、长时预测准确度不佳等问题；图神经网络针对路网拓扑连接关系引入了图结构，在考虑路网和交通流数据的时空相关性上具有明显优势；融合多因素网络充分考虑天气、道路、事故等内外部因素的影响，有效提升了交通流预测的实时性和鲁棒性；由于交通数据采集困难、外部因素影响难以量化、机器学习方法可解释性差等原因，交通流预测方法的改进受到了限制；未来应从交通信息有效挖掘和图卷积方法完善两方面入手，拓宽图结构在交通领域的应用和考虑非常态交通场景，进一步揭示交通数据的内在规律，开发更准确、高效的交通流预测方法，推动交通流预测在工业界的落地应用。

Review on machine learning-based traffic flow prediction methods

The research status and development trend of macro traffic flow prediction of designated road sections and regional road network at home and abroad were analyzed by literature review, expert interview, and experimental scenario construction. Local section traffic flow prediction methods were summarized, including traditional machine learning, recurrent neural networks, and hybrid models. The characteristics of convolutional neural networks, graph neural networks, and fusion multi-factor networks were discussed.The principles, advantages, limitations, and application scenarios of the methods were explained. The types of existing scenario traffic datasets and the mainstream traffic datasets at home and abroad were summarized from the perspectives of sampling periods and collecting methods. Analysis results show that recurrent neural networks can effectively obtain the historical laws of traffic data, but there are some problems such as gradient explosion, high computational complexity, and poor accuracy of long-time prediction. Graph neural networks introduce graph structures for road network topological connection relationships, which has obvious advantage in considering the spatiotemporal correlation of road network and traffic flow data. Fusion multi-factor methods fully consider the influence of internal and external factors such as weather, roads, and accidents, effectively improving the real-time performance and robustness of traffic flow prediction. The improvements of traffic flow prediction methods have limitations due to the difficult traffic data collection and external factor influence quantification, as well as the poor interpretability of machine learning methods. The future research should start from two aspects of starting the efficient mining of traffic information and the perfection of graph convolution methods, broaden the application of graph structures in the traffic field, and consider non-constant traffic scenarios. So as to further reveal the inherent laws of traffic data, develop more accurate and efficient traffic flow prediction methods, and promote the application of traffic flow prediction in industry.