Aiming at the problem that the time series characteristics of short-term car loading volume in railway freight station are difficult to be extracted due to its volatility and randomness, a combined prediction model STL-N-BEATS based on STL decomposition method and N-BEATS neural network model is proposed. Firstly, the original data is decomposed into trend series, periodic series and residual series by STL decomposition method. Then, the N-BEATS model is used to model each volume and reconstruct the prediction results. Finally, based on the 546 days historical car loading data in 4 freight stations of a railway transportation enterprise, the prediction performance of the proposed model is compared with the other 6 models. The results show that under the test set of station A, the predictions of the other 6 models have a certain lag, while the proposed model can better fit the real value curve, and the 3 indexes including the calculated symmetrical average absolute percentage error, average absolute error and root mean square error are the lowest. This is because the trend series and periodic series obtained by the proposed model after decomposing the time series characteristics dominate the prediction results, reducing the uncertainty and volatility of the overall data. When the prediction step is 3 and 7 days, under the prediction scenario of daily car loading capacity of stations B, C and D, and daily car loading capacity of different destinations and different commodity names of station D, the 3 indexes of the proposed model are still the lowest, which signifies its good prediction performance and generalization ability.
铁路货运量预测方面,目前国内外学者主要采用灰色模型、机器学习和深度学习等方法。肖金山等[1]采用灰色预测模型对我国2019—2020年全国铁路货运量需求进行预测;徐莉等[2]引入残差,对灰色模型进行修正,改善了预测结果。孙斌等[3]采用自回归积分滑动平均模型(Autoregressive Integrated Moving Average Model,ARIMA)对2020年全国铁路货运量进行预测。汤银英等[4]针对铁路月度货运量的趋势性与季节性特征,利用Holt-Winter模型对铁路月度货运量进行预测。王治[5]通过遗传算法确定支持向量机(Support Vector Machine,SVM)参数,利用少量历史数据预测了昆明市2002—2005年铁路货运量。梁宁等[6]将灰色关联分析法与SVM相结合,提升了货运量预测模型的泛化能力。谭雪等[7]采用门控循环单元(Gate Recurrent Unit,GRU)深度神经网络分别建立单步和多步模型,对铁路短期货运量进行预测并将预测结果与其他模型进行对比,证明其有效性。郭洪鹏等[8]使用双向长短时记忆(Bidirectional Long Short Term Memory,Bi-LSTM)网络预测了某铁路运输公司的月度货运量和日货运量。Liu等[9]利用Informer模型预测车站货运量,并通过不同模型的预测结果对比,证明该模型的预测性能更为优越。Hunt[10]使用ARIMA模型对爱沙尼亚的铁路货运量进行预测,其预测误差在10%左右。Milenkovic等[11]对阿尔卑斯—西巴尔干铁路货运走廊上4个边境口岸的进出口货运量进行预测,与文中其他模型相比LSTM模型的预测准确度更高。
上述文献主要从宏观角度对某一区域内的货运量预测进行研究,且大多在年、月等较长的时间尺度下进行预测,难以较好地对短期铁路日常工作计划的编制和调整进行指导,相比之下,货运站短期装车量的预测研究对于日常工作计划的制定和货运组织的调整更为有益。在此方面,余姣姣[12]首次使用SVM模型预测货运站装车量,但未解决相空间重构参数选择带来的模型不稳定问题,因此该模型的泛化能力较弱。张志文等[13]利用结合注意力机制(Attention Mechanism)的长短期记忆(Long Short Term Memory,LSTM)网络,研究了某一区域所有货运站的日装车量整体趋势,但未结合具体货运站实例讨论该方法的效果与性能。总体来看,多数针对货运站的短期装车量预测研究在建模时未考虑客户需求、车站计划、机车能力和装卸器具等因素导致日装车量存在一定的随机性和波动性,因此预测结果的准确度有限。
为了提高铁路货运短期装车量预测的准确性与模型的泛化能力,本文先提出货运站短期装车量预测的难点在于波动性和随机性两个特征的交织作用,再将基于Loess的时序分解方法(Seasonal-Trend Decomposition Procedure Based on Loess, STL)[16]与具有可解释性的神经网络预测模型(Neural Basis Expansion Analysis for Interpretable Time Series Forecasting,N-BEATS)相结合,提出一种铁路货运站短期装车量组合预测模型STL-N-BEATS(为便于行文,以下简称为“组合模型”)。该模型采用STL分解方法对原始数据进行分解,以期减少原始数据不同时序特征之间的干扰影响;借助神经网络对时间序列数据的强大学习能力,对分解后的数据序列分别进行预测,并通过线性叠加得到最终预测结果。
XIAOJinshan, HETao. Forecast of Railway Freight Volume Based on Improved Gray GM(1,1) Model [J]. Journal of Lanzhou Jiaotong University, 2021, 40 (3): 40-45. in Chinese
XULi, XUEFeng. Prediction of Rail Freight Volume of China Based on Secondary Residual Error Modification of GM(1,1) Model [J]. Journal of Transportation Engineering and Information, 2019, 17 (2): 44-50. in Chinese
SUNBin, CHENSiling, DULihui. Study on Prediction of Railway Freight Volume under Extreme Events Based on ARIMA Model [J]. Journal of East China Jiaotong University, 2021, 38 (2): 67-72. in Chinese
TANGYinying, LILong. Prediction Research about Monthly Traffic Volume Based on Holt-Winters Model [J]. Journal of Transportation Engineering and Information, 2017, 15 (2): 1-5, 13. in Chinese
WANGZhi. Prediction of Railway Freight Volume Based on Genetic Algorithm - Support Vector Machine [J]. Computer Simulation, 2010, 27 (12): 320-322, 348. in Chinese
LIANGNing, GENGLiyan, ZHANGZhanfu, et al. A Prediction Method of Railway Freight Volumes Using GRA and SVM-Mixed [J]. Journal of Transportation Systems Engineering and Information Technology, 2016, 16 (6): 94-99. in Chinese
TANXue, ZHANGXiaoqiang. GRU Deep Neural Network Based Short-Term Railway Freight Demand Forecasting [J]. Journal of the China Railway Society, 2020, 42 (12): 28-35. in Chinese
GUOHongpeng, LIUBin, XIAOYao. Short-Term Railway Freight Volume Prediction Based on Bi-LSTM Network [J]. Railway Freight Transport, 2022, 40 (2): 52-58. in Chinese
[17]
LIUJ Q, JINGY. Improving the Accuracy of Multi-Step Prediction of Railway Freight Volume Based on Informer Model [C]// 2024 4th International Conference on Consumer Electronics and Computer Engineering (ICCECE). Guangzhou, China. New York: IEEE Press, 2024: 444-448.
[18]
HUNTÜ. Forecasting of Railway Freight Volume: Approach of Estonian Railway to Arise Efficiency [J]. Transport, 2003, 18 (6): 255-258.
[19]
MILENKOVIĆM, GLIGORIĆM, BOJOVIĆN, et al. A Comparison between ARIMA, LSTM, ARIMA-LSTM and SSA for Cross-Border Rail Freight Traffic Forecasting: the Case of Alpine-Western Balkan Rail Freight Corridor [J]. Transportation Planning and Technology, 2024, 47 (1): 89-112.
[20]
余姣姣.基于SVM的铁路货运站装车数预测[J].物流技术,2019,38(3):38-42.
[21]
YUJiaojiao. Prediction on Car Loading Quantities of Railway Freight Stations Based on SVM [J]. Logistics Technology, 2019, 38 (3): 38-42. in Chinese
ZHANGZhiwen, LIUYongzhuang, ZHOUJin. Daily Car Loading Volume Prediction of Regional Freight Transport for China Energy Based on Attention Mechanism-LSTM [J]. Railway Transport and Economy, 2022, 44 (): 189-194. in Chinese
ZHAOMingwei, ZHANGWensheng. Short-Term Passenger Flow Prediction of Urban Rail Transit Based on IPSO-LSTM Combined Model [J]. Railway Transport and Economy, 2022, 44 (2): 123-130. in Chinese
HUANGBaojing, MAJun, YUYuanling. Prediction and Development Strategies of Xi’an China Railway Express Based on the SARIMA-RF Combination Model [J]. Journal of Chongqing Jiaotong University (Natural Science), 2024, 43 (10): 90-96. in Chinese
[28]
CLEVELANDR B, CLEVELANDW S. STL: a Seasonal-Trend Decomposition Procedure Based on Loess [J]. Journal of Official Statistics, 1990, 6 (1): 3-33.
WANGDewen, ZheLÜ. Multi-Step Prediction of Transformers Top Oil Temperature Based on STL and Attention Mechanism [J]. Electric Power Science and Engineering, 2022, 38 (11): 1-8. in Chinese
LEIQingwen, GAOPeiqiang, LIJianlin. A Monthly Runoff Forecast Model Combined Time Series Decomposition and CNN-LSTM [J]. Journal of Changjiang River Scientific Research Institute, 2023, 40 (6): 49-54. in Chinese
ZHAOJiandong, ZHUDan, LIUJiaxin. Metro Transfer Passenger Flow Prediction Based on STL-GRU [J]. Journal of South China University of Technology (Natural Science Edition), 2022, 50 (5): 22-31. in Chinese
[35]
陈光伟.铁路信息系统应用技术[M].北京:中国铁道出版社,2017.
[36]
CHENGuangwei. Application Technology of Railway Information System [M]. Beijing: China Railway Publishing House, 2017. in Chinese
[37]
ORESHKINB N, CARPOVD, CHAPADOSN, et al. N-BEATS: Neural Basis Expansion Analysis for Interpretable Time Series Forecasting [EB/OL]. arXiv: 1905. 10437 (2020-02-20) [2024-08-06].