RESOURCE LOAD PREDICTION OF CLOUD COMPUTING BASED ON STL-DEEPAR-HW COMPOSITE MODEL

With the vigorous development of informatization, the efficient management of a large number of cloud computing resources is an important problem in the field of operation and maintenance. Accurate load forecasting is the key technology to solve this problem. To solve this problem, a combined prediction model STL-DeepAR-HW based on seasonal and trend decomposition using loess STL, Holt-Winters model and deep autoregressive model DeepAR is proposed. The periodicity of data was extracted by fast Fourier transform and autocorrelation function, and the data was decomposed into trend items, seasonal items and remainder items by STL decomposition with the extracted optimal period. DeepAR and Holt-Winters were used to predict the trend items and seasonal items respectively, and finally the prediction results were combined. Experiments on AzurePublicDataset show that the combined model has higher accuracy and applicability in load forecasting than Transformer, Stacked-LSTM and Prophet models.