LLM-Guided Semantic Factorization Boosts Industrial Forecasting

Researchers have introduced Task-Semantic Field Factorization (TSF), an innovative framework that leverages large language models (LLMs) to enhance time-series forecasting and soft sensing in process industries. These industries often face challenges like scarce labeled data, frequent operational changes, and the high cost of model retraining. TSF addresses this by using LLMs offline to construct a "task-semantic field" from existing variable tables and process documentation, capturing the physical meanings and process roles of variables. During online training and inference, this semantic information is activated by the current numerical window, allowing it to participate in each prediction. This approach enables the model to adapt more effectively to different prediction targets and shifts in operating conditions. The framework significantly improves forecasting accuracy, reducing Mean Absolute Error (MAE) by an average of 6.4% across various complex industrial tasks, with some reductions reaching 25.5%. Crucially, TSF achieves these gains with minimal additional parameters and negligible online inference overhead, making it practical for deployment.