A DBN-Based Ensemble Method for Resource Usage Prediction in Clouds

Abstract

To accurately predict the dynamic changes on the usage of cloud computing resources so as to improve resource management efficiency, an ensemble method based on deep belief network (EMDBN) is proposed in this paper. Technically, it adopts the autoregressive model (AR) and the grayscale model (GM) as basic predictors. The results of the basic predictors are used together with historical data as training data for the DBN. Real data from the Google Cloud Data Center is used as experimental data. Back-propagation (BP) neural network, support vector regression (SVR), GM, AR and deep belief network (DBN) are chosen to carry out comparative experiments on cloud resource usage forecasting. Experimental results prove that EMDBN method can prominently enhance the precision in light of three current metrics for predict methods, i.e., mean absolute error (MAE), mean square error (MSE) and mean absolute percentage error (MAPE).