Abstract: Real-time prediction of load data collected from cloud platform monitoring helps in early identification of future system performance trends in cloud operations.However,load data typically lacks of clear periodicity or regularity and contains significant noise interference.Existing methods suffer from deficiencies in feature learning planning,relying on other load features and struggling to capture the momentum of load trends.To achieve accurate and efficient load data prediction,this paper proposes a cloud platform load data prediction method based on a spatiotemporal graph attention network.Firstly,an improved empirical wavelet transform is applied to perform time-frequency domain transformation on the load data,reducing noise interference and obtaining effectively decomposed modal features.To enhance the model’s capability in handling spikes and non-periodic characte-ristics,key performance factors tailored to the load data characteristics are designed using financial technical indicators.Additionally,modal features and key performance factors are reconstructed with the original sequence to build a graph learning layer.The graph attention network is then used to dynamically capture the relationships between the load sequences and features,and a bidirectional long short-term memory network focuses on temporal dependency information.Experimental validation is conducted on load datasets from Amazon and Alibaba Cloud,and the results show that,compared to the best baseline model,RMSE is reduced by 13.44%,36.90%,7.41%,and 14.93% respectively on four datasets.

Key words: Cloud platform, Load forecasting, Empirical wavelet transform, Financial technical indicators, Graph attention network, BiLSTM network