Combining physical laws and ANN for predicting energy consumption of data center cooling systems

Cooling systems within data centers are known for their substantial energy consumption. Predicting their energy usage typically involves two primary methodologies: constructing models grounded in physical principles and developing data-driven models. While physical models may lack broad applicability, artificial neural network (ANN) models often sacrifice interpretability. Striking a balance between accuracy and interpretability is a significant challenge in model development. In this study, we propose a novel approach that combines physical principles with ANN methodologies. By leveraging the strengths of both approaches, this combined model aims to enhance prediction accuracy while preserving interpretability and applicability. Experimental data specific to cooling systems were utilized to compare the predictive performance of the physical models, ANN models, and the combined model. Results demonstrate that the mean relative error (MRE) for the physical model was 7.95%, for the ANN model was 13.44%, and for the combined model was 6.54%. Additionally, the root mean square error (RMSE) values for the three models were 352.6, 258.3, and 181.9, respectively.