Performance Prediction of a Vertical Downward Supply Direct Expansion Cooling System for Large Spaces Through Field Experiments

Performance prediction of an air-cooled direct expansion (DX) vertical downward-supply cooling system applied to large spaces is a key element for achieving efficient control and energy savings. Recent studies have predominantly relied on complex artificial intelligence (AI)-based or high-dimensional models that require a large number of input variables to achieve high predictive accuracy. In contrast, limited research has focused on developing simple, interpretable, and practically applicable models based on field-measured data. To address this gap, the present study proposes a physically grounded multiple linear regression model with a minimal number of variables, which can be implemented in practice using only three standard sensors: indoor air temperature, outdoor air temperature, and airflow rate. Field data were refined through physical criteria derived from ASHRAE standards (steady-state operation and removal of outliers) and by identifying steady-state ranges using the Kernel Density Estimation (KDE) method. A total of 133,718 valid samples were used for analysis. The proposed model achieved a coefficient of determination (R²) of 0.93, a root mean square error (RMSE) of 2.86 kW, and mean absolute error (MAE) of 2.31 kW, corresponding to approximately ±6% deviation from measured cooling capacity. These results satisfy the typical accuracy criteria in the HVAC field (R² > 0.9, error < 10%) and confirm high predictive reliability despite the model’s simplicity. The achieved accuracy implies that the proposed model can be extended to field-level performance prediction and energy-efficient operation. Comparison with second-order polynomial and nonlinear (1/T_out) models showed only marginal improvement in accuracy. Consequently, the proposed three-variable regression model introduces a practical framework for performance prediction and control of DX-type cooling systems that integrates simplicity, physical interpretability, and field applicability.