PS5-27. Transformer-based Forecasting of Livestock Body Weight for Precision Feed Delivery.

Abstract Accurate forecasting of beef cattle's body weight can support data-driven feed delivery strategies and improve bunk management by aligning feed allocation with expected short-term growth. In this study, a transformer-based deep learning model (Informer) was used to forecast individual animal body weight over a 7-day horizon using historical weight time-series data. The Informer architecture employs ProbSparse self-attention to efficiently capture temporal dependencies in sequential data. Input sequences consisted of seven consecutive daily body weight observations used to predict the subsequent 7-day weight trajectory. There were a total of 1,883 animals (123,922 daily records) in the dataset; this dataset contains intake data of both drylot and grazing animals. Out of 1,883 animals 1,650 are in training dataset corresponding to 111, 529 daily records and 1,650 (12,392) animals were in test dataset. Data were partitioned at the animal level, with 90% of individuals used for training and 10% reserved for testing to ensure evaluation of unseen animals. Input features were standardized using a scaler fitted on the training dataset. The model was configured with a hidden dimension of 128, four attention heads, two encoder layers, one decoder layer, and a feed-forward dimension of 512. Training was performed for 100 epochs using the Adam optimizer (learning rate = 0.001) with a ReduceLROnPlateau scheduler. The model demonstrated strong predictive performance, achieving a training MSE of 63.37 and MAE of 3.73 kg (R² = 0.9899), and a test MSE of 34.25 with MAE of 3.44 kg (R² = 0.9950; peak R² = 0.9957). Seven-day forward forecasts across five animals projected a total gain of 11.51± 1.25 kg from the last recorded scale weight, equivalent to an average daily gain (ADG) of approximately 1.64 ± 0.18 kg/day. These results demonstrate that transformer-based time-series models can accurately capture short-term weight dynamics in livestock. Integrating such forecasts into feeding systems could enable proactive feed delivery adjustments and support slick bunk management by anticipating near-term growth and feed demand.