Abstract: Aiming at the problem that the load of mobile virtual machine changes greatly and it is difficult to be predicted accurately, this paper proposed an AR-LSTM-ED load forecasting model based on joint feature selection, which can carry out single-step and multi-step prediction of cloud host load. In this paper, the joint feature selection method was used to obtain other load series related to the target prediction load series, and the undecimated wavelet transform method suitable for online prediction was used to decompose the target prediction features into subsequences which could be easier to be predicted. Finally, these sequences and target prediction sequences were input into AR-LSTM-ED model. The model used the long-short term memory encoder-decoder network to predict the target load, which could capture the long-term dependence. The autoregressive model (AR) was further combined to predict the linear data in the load. We used the Google cloud computing data set to verify the algorithm. The experimental results show that the proposed method achieves better performance.

Key words: mobile cloud data center, LSTM neural network, feature selection, wavelet transform, load prediction