Related Experiment Video
Updated: Jun 25, 2026
13:27
Exploring the Effects of Atmospheric Forcings on Evaporation: Experimental Integration of the Atmospheric Boundary Layer and Shallow Subsurface
Published on: June 8, 2015
Incorporating dataset-level semantic priors into large language models for environmental time-series forecasting: a
1College of Internet and Big Data, Shenzhen Technology University, Shenzhen, 518118, China.
Scientific Reports
|June 16, 2026
Summary
This study introduces a semantic-prior-augmented large language model (LLM) for environmental forecasting. Integrating semantic data with numerical time series improves soil moisture and temperature predictions, especially in the mid-term.
Area of Science:
- Environmental science
- Data science
- Machine learning
Background:
- Environmental time-series forecasting faces challenges like non-stationarity and limited context.
- Large language models (LLMs) show promise but struggle to integrate global environmental characteristics beyond local data.
Purpose of the Study:
- To introduce a semantic-prior-augmented LLM (SP-LLM) framework for enhanced environmental time-series forecasting.
- To integrate numerical soil data with dataset-level semantic representations for improved contextual understanding.
Main Methods:
- Developed an SP-LLM framework combining numerical soil moisture and temperature time series with semantic priors.
- Constructed dataset-level semantic priors from training data to represent system behavior.
- Evaluated the framework on high-frequency greenhouse soil data across various forecasting horizons.
Main Results:
- Incorporating semantic priors reduced mid-term forecasting error by 4-8% MAE compared to numerical-only LLM models.
- Improved forecast stability for soil moisture and temperature across 6-12-hour horizons.
- Observed diminishing performance differences at longer horizons (24-48h), indicating system predictability limits.
Conclusions:
- Semantic representations effectively complement numerical data in environmental modeling.
- LLM-based approaches demonstrate horizon-dependent strengths and limitations in dynamic environmental systems.
- The SP-LLM framework offers a novel approach to enhance environmental forecasting accuracy and stability.
