Scholarlink

This paper proposes a novel risk analysis framework for the optimization of rolling stock management in rail freight shunting operations. We challenge the direct application of Machine Learning (ML) as input for operational decision-making by employing Risk Assessment strategies to evaluate how ML predictions affect the decision-making process. Our approach integrates the ML model’s performance metrics into a Mixed-Integer Linear Programming (MILP) model for shunting operation. A comparative analysis based on real data from the Luxembourgish rail freight company CFL Multimodal across various destinations reveals that a risk assessment approach provides superior performance compared to the direct use of the ML input, reducing the analyzed KPIs. This study demonstrates that the use of a risk assessment framework helps mitigate potential for operational inefficiencies and unfeasibility inherent in ML-dependent models.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Towards Safer Freight Rail Shunting: Integrating MILP and ML Classification Models in a Risk Management Framework

  • 章节

  • 发表日期：

    2026年

查看全文
  • Federico Bigi,
  • Tommaso Bosi,
  • Andrea D’Ariano,
  • Francesco Viti

摘要

This paper proposes a novel risk analysis framework for the optimization of rolling stock management in rail freight shunting operations. We challenge the direct application of Machine Learning (ML) as input for operational decision-making by employing Risk Assessment strategies to evaluate how ML predictions affect the decision-making process. Our approach integrates the ML model’s performance metrics into a Mixed-Integer Linear Programming (MILP) model for shunting operation. A comparative analysis based on real data from the Luxembourgish rail freight company CFL Multimodal across various destinations reveals that a risk assessment approach provides superior performance compared to the direct use of the ML input, reducing the analyzed KPIs. This study demonstrates that the use of a risk assessment framework helps mitigate potential for operational inefficiencies and unfeasibility inherent in ML-dependent models.