Cloud workload prediction by the DE-based nonstationary transformer model

Despite the revolution and convenience cloud computing has brought about, its huge energy usage is still a threat to environmental sustainability. Accurate cloud workload prediction enables cloud systems to manage and utilize resources more efficiently, which is a critical step in solving energy consumption problems. To address the issue of previous models' inability to capture non-stationary temporal features of cloud workload data, and the lack of predictive generalization for cloud workload with different patterns, a differential evolution based Non-stationary Transformer model is proposed to validate the feasibility of the model for cloud workload prediction problem in this paper. In the experimental part of the thesis, the model and the benchmark models are applied to the prediction task of cloud workloads in Google clusters with different load patterns. Through comparative experiments, the effectiveness and generalizability of differential evolution based Non-stationary Transformer model over cloud workload prediction task is demonstrated, and its predictive performance can meet or even exceed most of the benchmark models.