Simple-TimeLLM: A Simplified Quantization Approach to Rail Temperature Time-Series Forecasting with Large Language Models

Extreme weather events due to climate change are increasing the risk of railway track buckling, which can lead to delays and safety issues such as derailments. Accurate prediction of rail temperature is therefore crucial for safe train operations. While advancements in technology have enabled real-time monitoring of rail temperatures, the current machine learning models used by KORAIL face limitations in accuracy and generalization. This study introduces an innovative method utilizing pre-trained large language models (LLMs) for efficient rail temperature prediction. We employ a simple data mapping technique using binning and normalization to convert time-series data into token sequences suitable for LLMs, allowing training and inference without modifying the model architecture. By applying a non-autoregressive inference approach, we address error accumulation and facilitate faster inference. Our model, trained on five years of data from 200 sensors, outperformed various baseline models achieving an RMSE of 1.8539℃, MAE of 1.0681℃, and R² of 0.9794. It also demonstrated strong generalization in zero-shot predictions for sensors not included in training. This approach offers a straightforward and efficient way to apply LLMs to time-series data, contributing to safer railway operations and maintenance planning.