DRAFT-LLM: Retrieval-augmented frequency alignment with frozen large language models for cross-domain air-quality forecasting in enclosed sheep barns

Precise regulation of closed sheep barn environments is critical for animal welfare and production performance yet faces severe "cross-domain transfer" challenges. Variations in regional climates (e.g., extreme cold in Daqing vs. arid in Aksu), seasonal shifts, and dynamic ventilation/manure management induce strong non-stationarity and distribution shifts in pollutant data, rendering single-scenario models ineffective for new barns. Traditional approaches typically necessitate per-barn retraining, incurring high maintenance costs and hindering scalable deployment. To this end, we present DRAFT-LLM, a domain-aware time series forecasting framework for cross-barn, cross-season, and cross-region scenarios. This framework leverages a frozen Large Language Model (LLM) as a forecasting backbone, supported by retrieval-augmented context calibration and multi-scale frequency alignment. Specifically, we integrate RevIN and Patching at the input to adapt to non-stationary sequences, introduce FFT-based frequency soft prompts to capture periodic physical fingerprints, and construct an external Context Bank to retrieve similar historical conditions, mitigating distribution shifts. Concurrently, a goal-oriented asymmetric loss function is employed to enforce adherence to safety thresholds. Based on multi-source datasets spanning Heilongjiang and Xinjiang, we establish a rigorous "Leave-One-Domain-Out (LODO)" evaluation protocol. Comprehensive evaluations demonstrate that under a 60-step forecast horizon, our model achieves RMSE as low as 0.68 (NH₃), 2.43 (PM 2.5 ), and 33.36 (CO₂), reducing error by 15%–17% on average compared to the state-of-the-art TIME-LLM. In cold-start settings with only 10 target domain samples, NH₃ prediction error decreases further, exhibiting robust rapid calibration capabilities. Compared to 7B-scale models, DRAFT-LLM based on a GPT-2 backbone achieves comparable accuracy with only 1.2 GB memory usage and 12ms inference latency. Our method effectively resolves the engineering conflict between "high-precision forecasting" and "low-compute edge deployment," providing robust support for low-cost, transferable warning and control in cross-regional sheep barns.