一种基于局部加权学习的自适应交通流预测机制

An Adaptive Traffic Flow Prediction Mechanism Based on Locally Weighted Learning

结合交通流的特征,提出了一种自适应的交通流预测机制。首先根据训练数据特点,按三相交通流理论对交通状态进行分类,从每个分类对应的训练数据集内提取相应的最佳邻域。在基于局部线性回归模型的预测中,根据邻域中数据点所处状态分别选择相应局部模型进行预测,最终预测结果为各局部模型预测值的加权平均。根据各模型误差确定当前数据所处状态,增量加入训练数据集。基于真实交通数据的实验证实该方法能够有效提高预测的准确率。

The authors propose an adaptive neighborhood adjusting mechanism based on certain domain knowledge, namely three phrase traffic theory. The training dataset is divided into three subsets, in which corresponding neighborhood sizes are generated. In the estimation process, the authors make predictions on data at different traffic state with respective locally weighted learning models whose kernel bandwidths are different. The final prediction is a combination of all the previous predictions. Experiments based on real traffic data prove that the proposed method can improve the forecasting accuracy.