Power Consumption Prediction of Edge Servers Based on Mixed Features and Self-Attention Mechanism

The prediction of power consumption is the basis for studying energy consumption. In previous studies, the functional regression approach is difficult to adapt to current complex server architectures. Moreover, edge servers usually have limited data volume, and it is difficult to meet deep learning demand for large datasets. This paper propose a prediction model based on mixed features and self-attention mechanism, aiming to achieve good prediction performance of the model on small-scale datasets. Feature extraction, feature emphasis, and power consumption prediction are the three-stage workflow of the model. We propose a feature extension sub-layer in the feature extraction phase based on mixed features to reduce the difficulty of extracting temporal features and enrich feature information. The feature emphasis phase aims at selecting the important feature information, as well as further filtering the feature information. A series of comparative experiments show that this model is able to predict the best value of the effect higher than the other models by 26.67%. The ablation experiments verify that the proposed sub-layer can effectively improve the performance of the model.