基于交互式BP-UKF模型的短时交通流预测方法

A Method for Predicting Short-term Traffic Flow Based on Interactive IMM-BP-UKF Model

城市路网具有功能多样、组成复杂、交通量大、交叉口多等特点。优化短时交通流预测模型能够增加交通状态判别的精准度,有利于市民预知交通出行信息,为交通诱导措施的发布提供数据支持,免陷入拥堵困境。针对目前短时交通流预测模型优化过程中出现的模型适应性差,使用条件要求高、单一模型无法准确地描述交通流在不同时段内的变化规律等因素造成短时交通流预测精准度低的问题,采用Kohonen神经网络对交通流数据进行聚类分析,令聚类得到的不同交通流模式下的交通流数据作为训练多个不同BP神经网络模型的输入,将不同神经网络模型与无迹卡尔曼滤波结合组成多个交通滤波器,完成交通流量的非线性预测与在线校准,最后使用交互式方法融合各估计器预测结果得出综合交通流预测结果。仿真实例构建了多个估计器和1个基于该方法的联合估计器,将以上各个估计器用于某断面的流量预测中,验证了几类估计器的流量预测性能。试验结果表明:该方法搭建的联合估计器在各种交通模式下较单估计器的预测准确性均有所提高,且在交通流遭遇多因素影响下发生特征变化时表现出一定的自适应性;相比于传统系统预测模型,大大降低了对训练数据量的要求,取得了较为满意的短时交通流预测效果。

Urban road network has the characteristics of various functions, complex composition, large traffic volume and many intersections. The optimization of short-term traffic flow prediction model can increase the accuracy of traffic state discrimination, help the public to predict traffic travel information and provide data support for the release of traffic guidance measures to avoid congestion dilemma. At present, in the optimization process of short-term traffic flow prediction model, the low accuracy of short-term traffic flow prediction is mainly due to the poor adaptability of the model, high requirements for the using conditions and a single model cannot accurately describe the variation of traffic flow in different time periods. Aiming at this problem, the cluster analysis on the traffic flow data is conducted by using Kohonen neural network at first. Then, the traffic flow data in different traffic flow modes obtained by clustering are used as the input for training multiple different BP neural network models. Next, combining different neural network models with unscented Kalman filter to form multiple traffic filters to perform nonlinear prediction and online calibration of traffic volume. Finally, the prediction results of each estimator are fused by using an interactive method to obtain the prediction result of the integrated traffic flow. The simulation example constructed multiple estimators and a joint estimator based on this method, and the above estimators are used for predicting the traffic volume in a section to verify the performance of traffic prediction for several types of estimators. The experimental result shows that(1) the prediction accuracy of the joint estimator built by this method is improved in various traffic modes compared with the single estimator, and the joint estimator shows certain adaptability when the traffic flow encounters multiple influencing factors, the requirement of training data amount is greatly reduced compared with the traditional system prediction model, and the satisfactory short-term traffic flow prediction effect is obtained.