Incorporating dataset-level semantic priors into large language models for environmental time-series forecasting: a soil moisture and temperature case study

Environmental time-series forecasting is challenged by non-stationarity, limited observability, and the absence of explicit representations of system-level context. While recent large language model (LLM)-based approaches have shown promise for sequential prediction, their ability to incorporate global environmental characteristics beyond local numerical observations remains insufficiently explored. This study introduces a semantic-prior-augmented LLM (SP-LLM) framework that integrates numerical soil moisture and soil temperature time series with dataset-level semantic representations describing recurring patterns, temporal resolution, and overall system behavior. These dataset-level semantic priors are constructed solely from training data and provide contextual information not directly encoded in local observation windows. Using high-frequency greenhouse soil data as a case study, we evaluate the proposed approach across short-, mid-, and long-term forecasting horizons and benchmark it against linear models, deep learning architectures, and foundation time-series models. Results show that incorporating semantic priors reduces mid-term forecasting error by 4-8% in mean absolute error (MAE) compared with numerical-only LLM models and improves forecast stability across 6-12-h prediction horizons for both soil moisture and soil temperature. At longer horizons (24-48 h), performance differences between models diminish, indicating intrinsic predictability limits of the system. These findings demonstrate how semantic representations can complement numerical data in environmental modeling and provide insight into the horizon-dependent strengths and limitations of LLM-based approaches for dynamic environmental systems.