短时交通流预测中的特征选择算法研究

Feature Selection Algorithm in Short-time Traffic Flow Prediction

短时交通流预测是智能交通系统的重要基础,其精度直接影响到交通控制和诱导的效果.对于交通流预测中的非参数回归方法,其中一个重要的问题是状态向量的选取.本文提出基于ReliefF和Delta Test的特征选择算法来对特征向量进行选择.首先使用ReliefF算法根据特征和类别的相关性对状态向量进行快速初步筛选,加快算法的执行效率.接下来以Delta Test为性能指标,使用遗传算法对状态分量的权重进行进一步优选.最后通过基于实际数据的算例,对本文方法优选的状态向量与时间序列状态向量,简单时空关联向量进行了对比.结果表明,本文的方法在一般交通状态条件下和突变交通状态下都具有较好的性能.

The short-term traffic flow prediction is an important foundation of intelligent transportation system, and its performance significantly influences the effectiveness of traffic control and guidance. For the non- parametric regression method in traffic flow prediction, one of the most important problems is state vectors selection. This paper presents a feature selection algorithm based on ReliefF and Delta Test to select feature vectors. Firstly, ReliefF algorithm is used to filter the state vectors according to the correlation between features and classes to speed up the efficiency of the algorithm. Then using Delta Test as performance index, genetic algorithm is used to optimize the weight of state component. Finally, the state vector selected by this method is compared with the state vector of time series and the simple Spatial-temporal correlation vector. Numerical results show that the proposed method outperforms the other two usually used methods under both general and abrupt traffic conditions.