基于迁移学习的轨道交通特殊OD客流预测研究

OD Rail Transit Passenger Flow Forecast for Special Sections Based on Transfer Learning

客流预测一直是轨道交通运营公司关注的重点,由于受到运输能力的限制等因素影响,部分OD的实际客流数据与真实需求有偏差,出现异常或者样本缺失,从而造成总体样本量偏小,直接采用这些样本进行预测会明显影响预测精度,但通过还原样本值增加样本量难度太大。根据上述特点选择基于实例的迁移学习,先确定源域的对象和范围,从源域中选择合适的样本补充到总体样本中,共同组成最终的训练样本数据集,完成迁移学习。同时选择改进的Boost算法,通过误差调整样本权重,不断迭代,得到最终的预测模型。结果表明:基于实例的迁移学习结合改进Boost算法的预测精度要好于传统集成学习、ARIMA模型、多元回归模型,为轨道交通运营公司对特定OD的客流预测提供新的有益尝试。

The prediction of passenger traffic volume has always been the focus of the intercity railway operation companies.Due to the limitation of transportation capacity,the actual passenger traffic volume data of individual sections sometimes has a gap with the actual demand and there are abnormalities or missing samples.This will obviously affect the final accuracy if the samples are directly used for prediction.However,it is very difficult to restore biased sample data to the true value.This paper chose transfer learning based on cases to remove unqualified samples from the target domain and replace them with corresponding samples from the source domain to form the final training sample dataset.At the same time,the paper selected the improved Boost algorithm,adjusted the sample weight through the size of the error,and got the final model.The results show that the prediction accuracy of transfer learning based on cases combined with an improved Boost algorithm is better than that of traditional integrated learning models,ARIMA model, and multiple regression model. This study provides a new beneficial attempt for the city railway operation companies to predict the passenger traffic volume in some sections in the future.