New TSF framework cuts industrial forecasting errors by 25%

Industrial process forecasting relies on time-series models, but scarce labeled data and shifting operating regimes make traditional retraining costly. A new paper from researchers Zong, Jia, Ren, and He introduces Task-Semantic Field Factorization (TSF), an LLM-guided framework that leverages existing variable tables and process documents. Unlike methods that treat inputs as anonymous columns, TSF builds a task-semantic field offline using a large language model. During online training and inference, the current numerical window activates variable semantics, enabling adaptation to different prediction targets and operating shifts. The LLM is used only for offline construction, keeping online inference lightweight.

On multiple complex industrial forecasting and soft-sensing tasks, TSF reduces mean absolute error (MAE) by 6.4% on average, with the largest improvement reaching 25.5%. Remarkably, this requires adding only 1.8 to 3.0 thousand parameters and less than 0.008 milliseconds per step of additional compute. The framework works across different time-series backbones and semantic generators, making it highly practical for deployment. By turning existing process documents into measurable forecasting gains, TSF addresses a key pain point in industry: improving model accuracy without frequent retraining or heavy infrastructure upgrades. For professionals managing manufacturing, energy, or chemical processes, this offers a low-cost path to better soft sensing and predictive maintenance.