Abstract
With the rapid development of large-scale artificial intelligence models, the construction of data centers has been gradually accelerating. Data centers are equipped with numerous servers and network devices, and energy consumption is expected to keep increasing in the future. To improve thermal management, data center energy modeling has become a new challenge. Traditional numerical simulations are time-consuming and poor transferable, while data-driven machine learning models provide a feasible solution for data center modeling. Physics-informed neural networks embed prior physical knowledge into the loss function as soft constraints, guiding the modeling process to follow both data patterns and physical laws. To decouple multi-physics field modeling in data centers, we propose a dual-branch physics-informed neural network (Dual-PINN). Meanwhile, we design an adaptive loss re-weighting mechanism on the loss function based on the logistic function (Sigmoid) nonlinear transformation to balance the convergence speed among different loss terms. In the experiments, to comprehensively evaluate the proposed Dual-PINN, we choose the standard PINN, standard neural network, and random forest for comparison. Finally, we used MAE, RMSE, and the number of trainable parameters to measure model performance in terms of accuracy and complexity. The results demonstrate that Dual-PINN achieves the highest prediction accuracy with the fewest