VENTILATION RATE PREDICTION IN NATURALLY CFD-DRIVEN MACHINE LEARNING MODEL

Abstract

. In facility agriculture, ventilation is a fundamental factor, particularly natural ventilation, which is essential for improving crop productivity and conserving energy consumption. Computational fluid dynamics (CFD) has recently emerged as a key tool for quantitatively analyzing and predicting natural ventilation. However, CFD simulations are computationally demanding and resource-intensive when applied across diverse environmental conditions. In contrast, machine learning (ML) enables rapid and accurate predictions within its trained data range but involves significant effort to construct training datasets and lacks reliability in extrapolation scenarios. To overcome these limitations and integrate the advantages of both methods, this study developed an ML model using a CFD-generated training dataset covering the desired range of environmental parameters. Natural ventilation rates were determined using a CFD model based on the tracer gas decay (TGD) method for 27 locations within a greenhouse, considering variations in wind direction, wind speed, and vent opening condition. These CFD-derived ventilation rates were used as training data for ML models. Multiple regression, random forest, support vector regression, and deep neural network models were constructed, and their predictive performance was compared. To address the constraint of limited CFD simulation cases, the bootstrapping technique was employed to expand the dataset. The accuracy of the developed ML models was evaluated, demonstrating the feasibility of utilizing CFD-generated data to construct ML models for ventilation rate prediction. This approach highlights the potential for combining CFD and ML techniques to optimize natural ventilation in facility agriculture.