Forecasting sensor-derived animal-state indicators can provide forward-looking information for precision sheep farming, but sheep responses are shaped by their barn environment, previous state, and routine operations. We developed Graph-Enhanced Contextual Long-Range Forecasting for Sheep Farming (GCL-Sheep), a management-context-aware model for multi-step forecasting of active duration, rumination duration, feeding duration, intense exercise duration, and body temperature in Hu sheep. The study used monitoring records from 115 Hu sheep in two farms and three barns, covering monitoring campaigns from March 2024 to December 2025After domain screening and preprocessing, the data were organized into five farm–season–barn domains, containing approximately 325,000 usable hourly records and 304,000 supervised samples. Barn environmental records, individual physiological/behavioral measurements, and management-operation data were aligned to hourly sequences. GCL-Sheep combines Cross-Variable Graph Construction, hierarchical management-context prefixes, and long-context temporal modeling. For the representative in-domain active-duration forecasting task at the 12 h horizon (H=12), GCL-Sheep reduced the mean absolute error and root-mean-square error by 20.0% and 19.2%, respectively, compared with the second-best baseline, and improved the coefficient of determination by 0.079. In Leave-One-Domain-Out evaluation for active-duration forecasting at H=12, it achieved an average coefficient of determination of 0.792, and few-shot target-domain fine-tuning further improved accuracy. A 96 h historical window achieved the best balance between accuracy and temporal coverage. These results indicate promising retrospective multi-step forecasting performance and suggest that sensor-based animal-state forecasting may provide decision-support information for inspection scheduling and environmental management in sheep farms; however, welfare-threshold-based early-warning and intervention effects still require prospective field validation.
Wang et al. (Fri,) studied this question.