基于数据驱动的微观交通流建模研究

发布时间：2018-02-25 09:32

本文关键词： 微观交通流数据驱动机器学习跟驰模型换道模型 NGSIM数据集　出处：《西南交通大学》2017年硕士论文　论文类型：学位论文

【摘要】：跟驰模型和换道模型在交通安全评价、微观交通仿真、自巡航控制、自动驾驶等领域均有广泛的应用价值,它们也是道路微观交通流理论的核心内容。传统的微观交通流模型有一个共同不足就是它们均是基于数学公式和交通流理论建立的数学模型,这就导致此类模型很难有效地反映驾驶员的感知、思考、决策等一系列心理和生理活动的不一致性和不确定性。本文从数据本身的角度出发,利用机器学习相关算法特有优势,开展了基于数据驱动的微观交通流建模研究,来弥补上述不足,探索微观交通流模型研究的新方向。首先,针对跟驰行为展开研究,以线性组合预测为基础,融合基于动力学的跟驰模型对安全因素的可控性优点和基于机器学习的跟驰模型的强大自学习优点,通过改进最优加权法中的目标函数,来建立线性组合跟驰模型;然后,针对换道行为中的换道决策阶段,分别利用机器学习算法—BP神经网络、支持向量机、随机森林建立了数据驱动的自由换道决策模型,并利用归一化、主成分分析法对NGSIM数据进行预处理,然后用处理后的数据对模型进行训练和测试,验证模型的有效性。另外,基于随机森林独有的优势对影响换道决策行为的因素的重要性进行了分析;最后,针对换道行为中的换道执行阶段展开研究,首次提出基于BP神经网络建立换道执行模型,借助BP神经网络强大的自学能力和非线性拟合能力等优点,来弥补传统模型的不足。结果表明:组合跟驰模型的预测精度优于Gipps模型,且可通过调整组合跟驰模型考虑真实性和安全性的权重,来达到控制预测速度的真实性和安全性的目的;数据驱动的换道决策模型均有较高的精度;在换道决策阶段,驾驶员更多地关注在于本车与目标车道上车辆之间的距离,而不是相对速度,且相对目标车道上的前车来说,后车对驾驶员的影响更大一些;数据驱动的换道执行模型具有非常高的精度,用数据驱动的方法来建立换道执行模型是可行且有效的。
[Abstract]:The following model and the change model have extensive application value in the fields of traffic safety evaluation, microscopic traffic simulation, self-cruise control, autopilot and so on. They are also the core contents of the road microscopic traffic flow theory. A common shortcoming of the traditional microscopic traffic flow models is that they are all mathematical models based on mathematical formulas and traffic flow theories. This makes it difficult for such models to effectively reflect the inconsistency and uncertainty of a series of psychological and physical activities such as perception, thinking, decision making and so on. From the point of view of data itself, this paper makes use of the unique advantages of machine learning related algorithms. In order to make up for the above deficiencies, a data-driven research on microscopic traffic flow modeling is carried out to explore the new direction of microscopic traffic flow model. Firstly, the research on car-following behavior is carried out, which is based on linear combination prediction. Combining the controllability of the dynamic car-following model to the security factors and the powerful self-learning advantage of the machine-learning-based car-following model, a linear combinatorial car-following model is established by improving the objective function in the optimal weighting method. In this paper, a data-driven decision model is established by using machine learning algorithm (-BP) neural network, support vector machine (SVM) and random forest, and the decision model is normalized. The principal component analysis (PCA) is used to preprocess the NGSIM data, and then the model is trained and tested with the processed data to verify the validity of the model. Based on the unique advantages of random forests, the importance of factors influencing the decision making behavior is analyzed. Finally, a new model based on BP neural network is proposed for the first time. The advantages of BP neural network such as self-learning ability and nonlinear fitting ability are used to make up for the shortcomings of the traditional models. The results show that the prediction accuracy of the combined car-following model is better than that of the Gipps model. We can control the reliability and safety of prediction speed by adjusting the weight of combination and car-following model to control the authenticity and security of prediction speed. The driver pays more attention to the distance between the vehicle and the vehicle in the target lane rather than the relative speed, and the rear car has more influence on the driver than the front car in the target lane. The data driven switch execution model has a high precision, and it is feasible and effective to establish the data driven method.
【学位授予单位】：西南交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U491.112

【参考文献】