Estimating the air exchange rates in naturally ventilated cattle houses using Bayesian-optimized GBDT

It is challenging to estimate the air exchange rate (AER) dynamically in naturally ventilated livestock buildings such as dairy houses due to the influence of complex and variable outdoor environmental factors, large opening ratios, and the confusion of inflow and outflow at openings. This makes it difficult to efficiently regulate the opening ratio to meet the ventilation requirements in naturally ventilated livestock buildings. In this study, the air exchange rates of naturally ventilated cattle houses (NVCHs) in different seasons and opening ratios were obtained through field measurements and computational fluid dynamics (CFD) simulations. A fast and efficient machine learning framework was proposed and examined to predict AER based on the gradient boosting decision tree (GBDT) combined with Bayesian optimization. Compared with commonly used machine learning models such as multilayer perceptrons (MLPs) and support vector machines (SVMs), the proposed GBDT model has higher prediction accuracy and can avoid falling easily into local optima. Compared with the existing mechanical model based on the Bernoulli equation, the proposed GBDT model showed a slightly higher prediction than the mechanistic model and was much easier to use in AER estimation when inputting easily collected environmental factors in practical applications. Using Bayesian optimization could dramatically reduce the computing time when determining the optimal hyperparameter for establishing the GBDT model, dramatically saving on computing resources. Based on the Bayesian optimized GBDT model, the desirable opening ratio of the side curtain can be determined for automatically regulating the AER of cattle houses in future applications.