| dc.description.abstract | Spatial and temporal thermal non-uniformity in commercial broiler houses poses challenges for bird welfare, production efficiency, and environmental management. This dissertation investigated commercial broiler house thermal environment across three connected studies, each building on a high-density network of 77 wireless temperature sensors deployed at bird level in tunnel-ventilated commercial houses in the southeastern United States during spring and late summer growout periods.
The first study applied unsupervised machine learning (K-Means and ST-DBSCAN) to characterize thermal zones during two 49-day flocks. Broiler houses consistently developed 4-5 persistent thermal zones, with inter-cluster temperature differences increasing from 0.48-1.27°C at day 14 to 1.65-2.61°C by day 49 in late summer. K-Means clustering provided better performance metrics (SS: 0.37-0.45, DBI: 0.77-0.87, CHI: 46.57-77.19) through predefined number of clusters, while ST-DBSCAN, despite lower performance scores (SS: 0.18-0.37, DBI: 0.94-1.46, CHI: 21.69-42.50), proved more effective for identifying transient microclimatic anomalies or hotspots by classifying extreme readings as noise. Persistent thermal hotspot mapping revealed that the center and tunnel fan end are locations where temperatures consistently exceeded cluster averages.
The second study quantified indoor temperature variability across four houses over 49 growout days using coefficient of variation and Generalized Linear Mixed Model. Season, ventilation mode, bird age, and longitudinal position were all significant drivers of within-house thermal variability. Spring exhibited higher mean daily coefficient of variation (6.64%) than late summer (5.78%), and daily variability peaked at 13.57-14.14% by day 49 in spring. Ventilation mode had a significant effect on hourly temperature CV (p < 0.0001). In late summer, tunnel ventilation produced the highest geometric mean hourly CV (2.72%), significantly exceeding transitional (2.38%) and minimum (2.30%) modes (p ≤ 0.0005). Significant cross-sectional effects were detected in both seasons (p < 0.0001), indicating that temperature variability was spatially structured along the house length. In spring, higher CV was generally observed toward the pad end, in late summer, variability was lower overall and increased toward the fan end.
The third study applied supervised machine learning to predict indoor air temperature at discrete spatial locations. The Extra Trees Regressor achieved the highest accuracy (R2 = 0.86, RMSE = 0.95°C, MAPE = 2.86%), with outdoor air temperature (44.84%), growout day (15.62%), and setpoint temperature (14.09%) as the dominant predictors. The model generalized across all four houses (R2 > 0.84), and sensor optimization analysis showed that a reduced 22-sensor network could represent the full spatial temperature distribution with RMSE of 0.115°C, reducing sensor requirements by 71% relative to the 77-sensor reference network.
Together, these studies established a framework for thermal zone identification, variability characterization, and predictive modeling that supports spatial-oriented, data-driven thermal management in commercial broiler production. | en_US |
