Rail freight is vital for economic growth due to its efficiency and environmental benefits, but its lack of fixed schedules due to various delay factors poses challenges for accurate Expected Travel Time (ETT) predictions. This research addresses the need for real-time, accurate and dynamic ETT predictions crucial for maintaining efficient supply chains by developing a novel predictive model that leverages real-time data. The model ensembles Graph Convolutional Network-Long Short-Term Memory (GCN-LSTM) and Kalman Filters (KF) models to capture the complex spatiotemporal interactions and diverse traction behaviours within the freight train railway network. The methodology comprises three phases: modeling, schedule generation, and dynamic updating. In the modeling phase, historical train movement data is used to develop predictive models, with KF handling state-space representation and GCN-LSTM capturing spatial and temporal dependencies. These models are ensembled to enhance prediction accuracy. The schedule generation phase estimates travel times using the ensembled model, the dynamic updating phase refines predictions using real-time data, while congestion is assessed by clustering congested areas with Density-Based Spatial Clustering of Applications with Noise (DBSCAN) and propagating these clusters through KF. The proposed model is compared with different state-of-art predictive models. The methodology’s effectiveness was validated using real-world data from Indian Railway freight operations. The proposed model demonstrated superior accuracy, with Mean Absolute Percentage Error of 19.51%, while the moving average-based model which was previously being used by the Indian Railway had an error of 44.34%. This approach, implemented as a decision support system for Indian Railways’ daily operations, provides advanced planning solutions to manage the growing complexities of rail freight logistics effectively.

中文翻译：

---

用于铁路货运网络动态预期行驶时间估计的数据驱动预测模型：案例研究

铁路货运因其效率和环境效益而对经济增长至关重要，但由于各种延误因素而缺乏固定的时间表，这给准确的预期旅行时间 （ETT） 预测带来了挑战。这项研究通过开发一种利用实时数据的新型预测模型，解决了对维持高效供应链至关重要的实时、准确和动态 ETT 预测的需求。该模型集成图卷积网络长短期记忆 （GCN-LSTM） 和卡尔曼滤波器 （KF） 模型，以捕获货运列车铁路网络内复杂的时空交互和不同的牵引行为。该方法包括三个阶段：建模、计划生成和动态更新。在建模阶段，历史列车运动数据用于开发预测模型，KF 处理状态空间表示，GCN-LSTM 捕获空间和时间依赖性。这些模型经过集成以提高预测准确性。计划生成阶段使用集成模型估计旅行时间，动态更新阶段使用实时数据优化预测，而拥堵则通过使用基于密度的噪声应用程序空间聚类 （DBSCAN） 对拥堵区域进行聚类并通过 KF 传播这些集群来评估。将所提出的模型与不同的最先进的预测模型进行了比较。该方法的有效性使用来自 Indian Railway 货运运营的真实数据进行了验证。 所提出的模型表现出卓越的准确性，平均绝对百分比误差为 19.51%，而印度铁路以前使用的基于移动平均的模型的误差为 44.34%。这种方法作为印度铁路公司日常运营的决策支持系统实施，提供了先进的规划解决方案，以有效管理日益复杂的铁路货运物流。