A server load prediction method based on deep learning

A server load prediction method based on deep learning, comprising the following steps:

• (1) collecting a load usages series of the server in a past time and dividing it into a sample series and a label series;
• (2) extracting periodic information of the load usages series;
• (3) establishing a periodic information adaptive model based on encoder-decoder architecture;
• (4) inputting the sample series, the label series, and the periodic information into the model for training;
• (5) inputting the load usages series of the server in the current time segment into the trained model, the load usages series of the server in the future time segment can be predicted and output;
• wherein, a specific implementation method of step (1) is as follows:
• dividing the load usages series of the server in the past time into two segments, with the load usages series of the previous time segment as the sample series and the load usages series of the later time segment as the label series, and the data proportion of the previous time segment is greater than that of the later time segment,
• wherein, a specific implementation method of step (2) is as follows:
• for the load usages series of the server in the past time segment, an autocorrelation coefficient signal of the series is first calculated, and an abscissa value corresponding to the second peak of the autocorrelation coefficient signal is taken as a period length T, and the abscissa value represents the time interval length;
• then, taking the load usages series of the T time length before the label sequence as the periodic information.
• wherein, the periodic information adaptive model is comprised of an encoder module connected with a decoder module, the encoder module comprises a downsampling layer, a multi-headed self-attention mechanism layer, and a LSTM (Long Short Term Memory, Long short-term memory Network) layer L1, the decoder module comprises of a LSTM layer L2 and three attention mechanism layers A1˜A3,
• wherein, a specific implementation method of step (4) is as follows:
• first, dividing the sample series into two segments with the front segment series being longer than the back segment series, inputting the front segment series into the downsampling layer for downsampling, and inputting the downsampling result into the multi-headed self-attention mechanism layer;
• inputting the back segment series into the LSTM layer L1, inputting the hidden state of the LSTM layer L1 into the attention mechanism layer A1 as a query, and using the output of the multi-headed self-attention mechanism layer as the key and value pair of the attention mechanism layer A1, the output of the attention mechanism layer A1 is used as the initial hidden state of the LSTM layer L2, and the output of the LSTM layer L1 is used as the input of the LSTM layer L2;
• the decoder module needs to input the previous hidden state of the LSTM layer L2 into the attention mechanism layer A2 as a query during a decoding process, using the output of the multi-headed self-attention mechanism layer as the key and value pair of the attention mechanism layer A2, and using the output of the LSTM layer L2 and the periodic information together as the key and value pair of the attention mechanism layer A3, at the same time, using the back segment series of the sample series as a query of the attention mechanism layer A3, the output of the attention mechanism layer A3 is the final prediction result of the model;
• finally, the loss function between the prediction result and the label series is calculated. According to the loss function, the model parameters are backpropagated and updated by using the subgradient method, and the iterative update is continued accordingly until the loss function converges and reaches the minimum value or reaches the maximum number of iterations, that is, the training is completed;
• wherein, the expression of the loss function is as follows:
• Loss
• (
• X
• target
• ,
• X
• ˜
• pred
• )
• =
• max
• t
• ∈
• T
• (
• X
• target
• [
• t
• ]
• -
• X
• ˜
• pred
• [
• t
• ]
• )
• 2
• wherein, Loss(Xtarget, {tilde over (X)}pred) is a loss function between the Xtarget and {tilde over (X)}pred, Xtarget is the label series, {tilde over (X)}pred is the model prediction result, t is the time, Xtarget[t] is the load usages of the server at time t in the label series, {tilde over (X)}pred[t] is the load usages of the server at time t in the model prediction result.