基于模型偏差学习的交通信号自适应优化方法

A model bias learning approach for adaptive traffic signal control

城市交通信号配时优化是保证交通系统整体运行效率的前提条件。传统基于模型的信号控制方法往往基于历史统计数据,并且底层交通流模型不可避免地存在模型偏差,影响控制方案的科学性及其实际应用效果。考虑模型与数据融合驱动,提出一种基于模型偏差学习的交通信号控制自适应优化方法,以最小化路网行程时间为目标,建立信号控制最优化模型。首先,针对交通流预测模型的模型偏差,引入偏差函数表示预测模型与实际交通流状态间的误差;其次,建立基于径向基函数(RBF)神经网络的偏差函数,结合实际交通流数据对模型偏差进行学习,提高偏差函数的拟合效果;在此基础上,提出考虑模型偏差信息的自适应优化方法,以提高信号控制方法的控制性能。以小型测试路网与实际路网为例进行算例分析,基于SUMO仿真对本文提出的模型偏差学习方法进行验证,考虑不同交通流量条件,与固定配时、模型预测控制进行比较,分析其控制性能指标。结果表明,模型与数据融合驱动的信号控制自适应优化方法能有效提高预测模型的准确性,降低路网行程时间,与基于模型的控制和模型预测控制方法相比路网累积行程时间平均减少了38.3%和25.6%,提升了路网的实际运行效果。最后在宣城实际路网的仿真验证了方法的有效性。

Urban traffic signal control is an indispensable part of traffic management,which affects the traffic operation of the whole system.Traditional model-based control cannot effectively make use of the real-time traffic data.Moreover,the inherent prediction errors between model and reality usually lead to unreasonable signal timing plans.This paper proposed a model bias learning approach for adaptive optimization of traffic signal control that integrates the model-based method with the data-driven method.First,aiming at the deviation of the traffic flow prediction model,a model bias function was introduced to represent the error between the prediction model and the actual traffic flow state.Second,the model bias function was formulated based on the radial basis function(RBF)neural network,and the model deviation was learnt by combining the actual traffic flow data to improve the fitting effect of the deviation function.Furthermore,an adaptive optimization method considering the model bias information was proposed to improve the signal control performance.By taking the small test road network and the actual road network as examples,the model deviation learning method proposed in this paper was verified based on SUMO simulation.Considering different traffic flow conditions,the control performance indices were analyzed by comparing with fixed timing and model predictive control.Numerical results show that our proposed model bias learning approach can effectively improve the accuracy of the prediction model and reduce the total time spend of the system.Compared with the model-based control method and the model predictive control method,our proposed method reduces the cumulative network total travel time by 38.3%and 25.6%on average,respectively,improving the performance of the optimal control.Finally,the effectiveness of the method was verified on the Xuancheng network.