Accurate prediction of overall trends and peak loads in container cloud resource requests is crucial for improving resource utilization and optimizing allocation strategies. To address the limitations of existing prediction methods, such as insufficient accuracy of single models and inadequate capture of long-term temporal features, a hybrid cloud resource prediction model was proposed, which integrated polar lights optimization (PLO), variational mode decomposition (VMD), temporal convolutional network (TCN), and Informer (PLO-VMD-TCN-Informer). In the combined model, structures such as dilated convolutions were utilized to effectively extract local time patterns and enhance the model’s ability to capture detailed features. VMD was employed to improve the modeling capability for non-stationary and nonlinear time series, while PLO was applied to dynamically optimize the parameters of VMD and Informer, thereby increasing the level of modeling automation. Experiments based on the public Alibaba dataset show that the proposed model outperforms mainstream benchmark models across all key metrics. Specifically, the root mean square error (RMSE) is reduced by 24.03% to 28.00%, and the mean absolute percentage error (MAPE) is reduced by 22.37% to 31.59%, validating the effectiveness of the proposed model.