A Hyperparameter Optimization Algorithm for the LSTM Temperature Prediction Model in Data Center

Abstract

As the main tool to realize data mining and efficient knowledge acquisition in the era of big data, machine learning is widely used in data center energy-saving research. The temperature prediction model based on machine learning predicts the state of the data center according to the upcoming tasks. It can adjust the refrigeration equipment in advance to avoid temperature regulation lag and set the air conditioning temperature according to the actual demand to avoid excessive refrigeration. Task scheduling and migration algorithm based on temperature prediction can effectively avoid hot spots. However, the choice of hyperparameter of machine learning model has a great impact on its performance. In this study, a hyperparameter optimization algorithm based on MLP is proposed. On the basis of trying certain hyperparameters, the MLP model is used to predict the value of all hyperparameters’ space, and then, a certain number of high-quality hyperparameters are selected to train the model repeatedly. In each iteration, the amount of training data decreases gradually, while the accuracy of the model improves rapidly, and finally, the appropriate hyperparameter are obtained. We use the idea of mutation in the genetic algorithm to improve the probability of high-quality solutions and the loss function weighting method to select the solution with the best stability. Experiments are carried out on two representative machine learning models, LSTM and Random Forest, and compared with the standard Gaussian Bayes and Random Search method. The results show that the method proposed in this study can obtain high-precision and high-stability hyperparameter through one run and can greatly improve the operation efficiency. This algorithm is not only effective for LSTM but also suitable for other machine learning models.