基于注意力机制-LSTM的国家能源集团铁路区域货运日装车量预测研究

Daily Car Loading Volume Prediction of Regional Freight Transport for China Energy Based on Attention Mechanism–LSTM

在铁路货运日常运营中,货运装车和卸车作业是重要业务环节,预测不同时间粒度下的装车量和卸车量对掌握铁路货运工作的基本情况和变化趋势至关重要。但货运装车量不仅受货运订单的需求影响,也和区域内计划、车站、机车调度等环环相扣,具有复杂性。研究对国家能源集团铁路调度信息系统内业务数据进行统计分析和挖掘,提取与日装车量相关的其他特征,在此基础上首先比较传统的时间序列模型ARIMA和循环神经网络RNN中的LSTM对日装车量的数据拟合与预测效果,然后探讨单变量和多变量下LSTM模型以及带有Attention层的LSTM对日装车量的预测效果,实验结果不仅展现了时序神经网络在铁路货运各类时序预测任务中的优势,还验证了业务数据挖掘和特征工程对日装车量预测精确度提升的作用。

Loading and unloading of freight trains are pivotal links in the daily operation of freight trains.Prediction of loading and unloading volumes under various time granularities is important for understanding the basic information and trends of railway freight work.However,the loading volume of freight trains is not only affected by freight orders but also widely correlated with procedures such as regional planning,stations,and locomotive dispatching.Therefore,it is a complicated process.This paper conducted the statistical analysis and data mining based on business data in the China Energy railway dispatching information system and extracted features strongly correlated with daily loading volume.On this basis,this paper then compared the data fitting and prediction effects of traditional time-series model ARIMA and LSTM in recurrent neural network(RNN)and explored the prediction effects of the LSTM model and LSTM with an Attention layer under a single variable and multiple variables.As a result,the experimental results show the advantage of the time-series neural network in time-series prediction tasks of railway freight and prove that business data mining and feature engineering can improve the accuracy of daily loading volume prediction.