Scientists Propose TSF Method: LLM Improves Industrial Process Forecasting

Researchers published a framework on arXiv called Task-Semantic Field Factorization (TSF) — a method in which a large language model (LLM) pre-labels industrial data with semantic context, while the time series forecasting model itself continues to work in normal mode.

What Problem TSF Solves

In process industry, time series forecasting and "soft sensing" are needed to assess quality indicators that are difficult to measure online directly. The problem is that labeled data is scarce, operating modes change frequently, and retraining the model or rebuilding the pipeline for each new scenario is expensive. Meanwhile, companies typically already have variable tables and process regulations — with names, units of measurement, and physical meaning of each parameter — but standard time series models usually treat input data as unnamed numeric columns, not using this textual documentation.

How Much Accuracy Improved

TSF builds a "semantic field" from process regulations and variable descriptions before training, using the LLM only offline — at the preparation stage. During training and inference, the current numeric window activates the corresponding semantics for variables, so semantic information participates in each forecast.

• Average forecast error (MAE) decreases by 6.4% in improved scenarios
• In the best cases, MAE reduction reaches 25.5%
• The method adds only about 1.8–3.0 thousand parameters
• Additional inference latency is less than 0.008 ms per step

What This Means

The authors show that industrial process documentation, which typically just sits in company archives, can be turned into measurable forecast accuracy gains — with almost no increase in computational cost. This makes the approach suitable for easy integration into already running industrial models, not just for new pilot projects.

Frequently Asked Questions

What is soft sensing?

It is a technique for assessing process quality indicators that are difficult or expensive to measure directly online — instead of a direct sensor, indirect data and a model are used to predict the needed parameter. TSF was tested on such tasks, along with time series forecasting.

How heavy is TSF to embed in production?

The method adds only 1.8–3 thousand parameters and less than 0.008 ms inference latency per step, so it can be embedded in already deployed forecasting models without significant infrastructure changes.