Comparative Evaluation of Machine Learning and Temporal Models for Data Center Power Prediction

This study presents a comparative evaluation of machine learning and temporal models for predicting power consumption in data centers. Using real-world Google cluster trace data comprising 8,927 temporally ordered samples, we investigate the effectiveness of multiple predictive approaches, including naive temporal baselines, linear regression, Random Forest, and classical time-series models such as AR(1) and ARIMA. The dataset is constructed using aggregated 5-minute intervals with features including CPU utilization, lagged CPU values, lagged power, and derived feature transformations. A time-based train-test split is employed to preserve temporal dependencies and avoid data leakage. Model performance is evaluated using Root Mean Squared Error (RMSE). Experimental results show that a simple naive temporal model achieves the lowest prediction error, outperforming both machine learning and classical time-series approaches. Linear regression demonstrates comparable performance, while Random Forest slightly underperforms. Traditional time-series models and CPU-only models exhibit significantly higher error, indicating limited predictive capability. These findings suggest that temporal stability plays a dominant role in data center power behavior, particularly in low-utilization environments, and that workload features alone provide limited predictive value. The study highlights the importance of baseline temporal patterns over model complexity for accurate power prediction.