<p>The demand for cloud computing services has increased dramatically. With the promotion of global low-carbon policies, increasing energy savings and efficiency in cloud computing services is important. By improving load prediction capability, reasonable allocation of cloud service management resources can be effectively realized. However, it is difficult to effectively extract features, and the accuracy of load prediction is poor due to large fluctuations and irregular changes in the cloud resource load. Thus, in this study, we propose a parallel cloud resource load prediction model, PCL-RC, that is based on feature optimization and focuses on feature extraction optimization and load forecasting. To address the problem of nonlinear load data feature extraction, a feature extraction optimization method that is based on combining an improved random forest method and complete ensemble empirical modal decomposition with adaptive noise is proposed to realize regular decomposition and feature extraction from fluctuating data. To address the issues of increased data volume due to decomposition and low prediction accuracy due to difficulty in extracting hidden features, a cloud resource load forecasting method based on an improved lightweight attention mechanism long short-term memory network is proposed. Experiments are conducted on data from the AliCloud platform. The proposed model outperforms the AR, SVR, HAR, Informer, Transform, VMDSE-Tformer and XGBoost models and has improved prediction accuracy.</p>