Feature-driven optimization for growth and mortality prevention in poultry farms

The poultry sector faces ongoing challenges in reducing mortality and improving growth performance under variable environmental and operational conditions. To address this, we developed and implemented a feature-driven optimization model to predict two key indicators: mortality and average weight (AvgWeight), based on five input variables-day, temperature, humidity, feed consumption, and water consumption. An 88-day dataset from over 20,000 Taiwan native broilers was preprocessed through outlier removal, normalization, and interpolation. Five machine learning models-Random Forest, Gradient Boosting Machine, Support Vector Machine, Linear Regression, and Neural Network (NN)-were initially evaluated. The baseline NN demonstrated superior multi-output accuracy and was further refined into three variants. Among them, the Ensemble NN-comprising five parallel networks-achieved RMSEs of 0.45 (Mortality) and 0.02 (AvgWeight), with Coefficient of Variation of RMSE values of 3.42 % and 1.62 %, respectively. Feature importance and sensitivity analyses identified "Day" as the most influential predictor for Mortality (importance: 20.391; sensitivity: 42.513), followed by feed (12.785; 13.285) and water (11.426; 13.648) consumption, while environmental variables had less impact under stable housing. Integrated into a MATLAB-based application, this intelligent system enables "what-if" scenario-offering practical decision support for poultry managers. By bridging traditional livestock management with deep learning-based soft sensors, this study contributes a scalable and accurate tool for advancing precision agriculture in poultry production.