Yu, Wanju (Xi'an Jiaotong University), Li, Donghe (Xi'an Jiaotong University), Yang, Ye (Shanghai DianJi University), Zou, Yanbin (Xi'an Jiaotong University), Wang, Junyi (Xi'an Jiaotong University), Chen, Shitao (Xi'an Jiaotong University, Xi'an, China)

Large Language Model Guided Gear-Shifting Optimization for Energy-Saving Rail Operations

Scheduled for presentation during the Invited Session "S16a-Control, Communication and Emerging Technologies in Smart Rail Systems" (TH-LM-T16), Thursday, November 20, 2025, 11:30−11:50, Southport 1

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 18, 2025

Abstract

Energy-efficient operation of electric rail transit is crucial for green urban transportation, but faces challenges due to fixed timetables, strict speed constraints, and complex track characteristics. Existing methods, such as reinforcement learning or heuristic optimization, often suffer from poor generalization across tracks or high computational costs. This paper proposes a two-stage hybrid optimization framework, termed LLM-PSO-NM. First, a refined physical model of the electric rail locomotive is constructed, including track modeling, dynamic modeling, and energy consumption evaluation, providing an accurate physical simulation environment for the optimization process. In the first stage of the framework, a pretrained Large Language Model (LLM) interprets track parameters, travel time constraints, and speed limits to generate physically feasible initial gear-shifting strategies, offering high quality starting solutions. In the second stage, a hybrid optimizer that combines Particle Swarm Optimization (PSO) and the Nelder–Mead (NM) method refines the initial solutions based on energy consumption feedback from the physical modeling module, aiming to minimize the train’s traction energy consumption. This division of labor—using LLMs for initial solution generation and PSO-NM for fine-tuning reduces search complexity and accelerates convergence. Experiments on four real-world railway sections (over 150 km) show that the proposed method reduces energy consumption by up to 2.96% compared to PSO-based optimization and by up to 1.59% compared to PPO-based optimization. Furthermore, it shortens optimization time by approximately 28.86% compared to PSO and by approximately 28.83% compared to PPO. It also demonstrates strong cross-track transferability without retraining.