铁道科学与工程学报2025，Vol.22Issue(8)：3425-3435,11.DOI:10.19713/j.cnki.43-1423/u.T20241603
基于PG-ConvLSTM网络的铁路运用车分配预测方法
A prediction method for rail serviceable cars allocation based on PG-ConvLSTM network
摘要
Abstract
The efficient and rational determination of the rail serviceable cars allocation scheme for future monthly planning periods is crucial in satisfying freight market demands and reducing freight cars allocation costs.Targeting the spatiotemporal features of the rail serviceable cars allocation scheme,a prediction method for rail serviceable cars allocation called PG-ConvLSTM was proposed,which combined convolutional long short-term memory(ConvLSTM)neural network with physics-guided(PG)idea.Initially,through the identification of key influencing factors,three key influencing factors were screened out,including freight workload,freight organization level,and service attributes of freight cars.These factors were then constructed as model inputs using the sliding window division method.Subsequently,a prediction model for rail serviceable cars allocation based on the PG-ConvLSTM network was constructed.The model utilized the ConvLSTM neural network as the primary framework and designed physical inconsistencies based on a priori rules of rail serviceable cars allocation a physics-guided loss function.Additionally,the hyperband algorithm used the Hyperband algorithm to optimize two hyperparameters of the model:the number of network layers and the weight assigned to physical inconsistencies.Finally,by using MAE,RMSE and MAPE as evaluation metrics,and selecting the BP,CNN-LSTM,CNN-GRU,and ConvLSTM network models for comparison,a case analysis was conducted using actual data on rail serviceable cars allocation prediction.The results demonstrate that among the neural network models considered,the PG-ConvLSTM model exhibits superior performance with the MAE of 0.002 8,RMSE of 0.003 4,and MAPE of 7.22%.The exceptional predictive capability of the PG-ConvLSTM model can be attributed to its synchronous extraction mechanism for spatiotemporal correlation features,which effectively mitigates the loss of key information after the spatiotemporal features are convolved and then input into the recurrent neural network.Additionally,the physics-guided loss function positively contributes to enhancing prediction accuracy.The PG-ConvLSTM model can efficiently and accurately predict rail serviceable car allocation schemes,thereby providing references for formulating monthly planning period allocation strategies in practical operations.关键词
铁路运输/铁路运用车/分配预测/卷积长短期记忆神经网络/物理引导Key words
railway transport/rail serviceable car/allocation prediction/ConvLSTM/physics-guided分类
交通工程引用本文复制引用
龙泽雨,何世伟,王攸妙,吴艺迪..基于PG-ConvLSTM网络的铁路运用车分配预测方法[J].铁道科学与工程学报,2025,22(8):3425-3435,11.基金项目
中央高校基本科研业务费专项资金资助项目(2024JBZX038) （2024JBZX038）
中国国家铁路集团有限公司科技研究开发计划课题(N2024X022) （N2024X022）
成都局集团公司科技研究开发计划课题(CX23089) （CX23089）
