Feasibility of using neural networks for real-time prediction of poultry deep body temperature responses to stressful changes in ambient temperature

In this article, the feasibility of using artificial neural network (ANN) m odels for predicting deep body temperature (DBT) responses of broilers to s tressful step changes in ambient temperature was determined. Experiments we re carried out using three different birds exposed to five AT schedules. DB T responses were measured using telemetric sensors, Various ANN architectur es were tested and the Elman-Jordan was determined to be most suitable. A f actor analysis was conducted to determine input variables most appropriate for the prediction. Although relative humidity (RH) was maintained almost c onstant, including it as art input to the network led to improved predictio ns, The ability of the developed models to predict DBT responses to AT sche dules not used in training and/or responses from a bird not used in trainin g was examined. The models performed reasonably well when predicting respon ses of a different bird to AT schedules used in training, The models perfor med well when predicting responses of a bird used in training to new AT sch edules. However predictions of the models were less accurate when dealing w ith a different AT schedule oa a different bird. This latter result is not surprising considering that the network had to adapt to two new conditions with training based on a limited data set, Using a larger data set with mor e birds and more AT schedules would likely lead to improved DBT predictions , Results of this study indicate that neural networks could potentially be used for predicting the impact of heat stress conditions on bird physiology .