LLMs Enhance Industrial Forecasting by Integrating Semantic Process Knowledge.

Industrial processes often struggle with time-series forecasting and soft sensing due to scarce labeled data, frequent regime changes, and the high cost of retraining models. While process documents contain valuable semantic information about variables and their roles, standard time-series models typically treat inputs as anonymous numerical data, failing to leverage this context. This paper introduces Task-Semantic Field Factorization (TSF), a framework designed to bridge this gap. TSF employs an LLM offline to construct a comprehensive "task-semantic field" from existing task protocols and variable documentation. This semantic field is then integrated into conventional time-series backbones during training and inference. By activating variable semantics within each numerical window, TSF ensures that semantic information actively participates in every prediction, enhancing adaptability to different prediction targets and operating shifts. Experiments across various industrial forecasting tasks show TSF significantly reduces Mean Absolute Error (MAE) with minimal additional parameters and negligible online inference overhead, effectively turning existing process documents into measurable performance gains.