Research On Energy Consumption Feature Extraction And Explainable Energy Consumption Prediction Methods Of Cloud Data Center

With the advent of the "new infrastructure" era,the scale of data centers has become larger and larger,requiring data center energy consumption prediction and evaluation models to have a high degree of accuracy and interpretability at the same time.In addition,according to service strategies such as data center container pricing and security operation and maintenance strategies,the daily maintenance requirements of servers are mainly divided into physical machine-level hardware security maintenance and container-level workflow container scheduling.Therefore,this paper studies how to construct high-precision and highly interpretable energy consumption models at these two levels.Different characteristic data collection methods and interpretable energy consumption models are designed for the two levels..At the physical machine level,the number of features of the energy consumption model is small and fixed.This paper proposes an interpretable framework for physical machine hardware energy consumption,and uses an interpretability method based on the contribution of common features to interpret the model that has been built.Compared with the traditional feature contribution value interpretation method,this method achieves a better interpretability effect of the physical machinelevel energy consumption model.At the same time,because the method is a post-interpretability method,it retains the high accuracy of the original model.At the container level,the number of features of the energy consumption model is large and constantly changing.This paper proposes an interpretable framework for energy consumption based on container characteristics.And on this basis,the hierarchical neural network model is used to decompose the energy consumption of the physical machine into the sum of the energy consumption of the container,and then the energy consumption of each container is decomposed into the energy consumption of the resource allocated by the container.By analyzing the correlation between the energy consumption obtained by the decomposition of the container level and the resource usage of the container,it is verified that the method is highly interpretable at the container level.