Temperature prediction in data center combining with deep neural network

The stable temperature of the computer rooms in data centers (DCs) is crucial for the servers' safety operation. The increase in server temperature can lead to poor operational performance and even malfunctions due to high temperature. Therefore, accurate temperature prediction becomes particularly important. In this paper, a datadriven server temperature prediction model is constructed by combining numerical methods with deep learning prediction algorithms. Firstly, the numerical model of the computer room with servers configured is built, where the heat elements such as the CPU are equivalent to the integrated heat source (IHS). The Computational Fluid Dynamic (CFD) method is utilized for the numerical calculation of the temperature and velocity fields in the DC computer room. Besides, the correctness of the numerical model is verified through experiments. The experimental results show that the temperature Mean Absolute Error predicted by the numerical model does not exceed 0.83, and the temperature calculation accuracy reaches 99.7%. Then, the more efficient Deep Neural Network (DNN) is applied to train the temperature data generated by the CFD numerical model. Besides, an Attention Mechanism has been added to facilitate the capture of important feature information during the training process. Meanwhile, searching for optimal hyperparameters of the DNN model by Bayesian Optimization, and the Attention-DNN-Bayesian model is established for temperature prediction. The mean absolute error of the DNN model is 3.51, while the ADB model is 0.65, which reduces by 81.48%.