关键词: 铁路运营数据, 列车运行时间, 预测模型, Box-Cox转换, 决策树, 装箱梯度提升树(HGBT)

Abstract:

In order to accurately predict train running time while considering requirement on accuracy and timeliness of its operation, a time prediction model was established. Firstly, data of railway timetable were extracted and analyzed to determine possible influencing factors of its running time, and Box-Cox transformation was used to normalize operation data considering their non-normality. Then, the model's input features and hyper-parameters were optimized based on decision tree and grid-search algorithm, respectively, and its performance was improved. Finally, a train running time prediction model was established by adopting HGBT based on optimized input features and hyper-parameters, and operation data of a Chinese and European railway were utilized to evaluate works at each stage. The results show that Box-Cox transformation can significantly improve data normality and goodness-of-fit of the prediction model, while grid-search algorithm can simultaneously improve the model's efficiency and accuracy. Compared with other commonly used machine learning algorithms for train running time prediction, the proposed HGBT model features higher accuracy and efficiency.

Key words: railway operation data, train running time, prediction model, Box-Cox transformation, decision tree, histogram-based gradient boosting tree (HGBT)