Dynamic calibration of broiler apparent temperature using real-time panting rate and machine learning: a pilot study

To address the need for precise thermal environment assessment in intensive broiler farming, this study proposes a dynamic apparent temperature (AT) optimization model that incorporates real-time biological feedback, overcoming the limitations of static traditional AT models. A baseline “AT-P” curve was established using the flock's minute-level panting rate (P) as a core biological feedback indicator. This curve is based on 7,138 records collected from a commercial farm, which include temperature, humidity, air velocity, age, and synchronously captured panting rate data. This curve maps observed panting to an “equivalent AT” representing the actual thermal load, with the deviation from traditional AT defined as the systematic correction residual. An Enhanced Attention Gradient Boosting Machine (EAG) ensemble model is then introduced to dynamically predict this residual, taking multi-source environmental features, traditional AT, and real-time panting rate as inputs, and outputting a corrected, optimized AT. Experimental results demonstrate that the EAG model achieves optimal performance in residual prediction, with an R² of 0.8329 and a mean absolute error (MAE) of 0.6503, significantly outperforming single base models and other mainstream algorithms. By integrating a static physical model with dynamic group behavioral feedback for online self-calibration, this study provides a methodological foundation for developing an animal-centric “perception-response-optimization” intelligent environmental control system in smart farming.