Transformer Encoder-Based Workload Prediction for Peak-Aware Energy Efficiency in Edge Computing

Energy-efficient management in edge computing systems is a critical challenge due to highly dynamic workloads and limited computational resources. Unpredictable peaks in CPU and memory usage can lead to excessive power consumption, thermal stress, and reduced hardware longevity. This study proposes a Transformer encoder-based time series regression model for workload prediction, designed to anticipate future demand and enable proactive, energy-aware decision-making. By learning temporal patterns in historical workload traces, the model effectively captures fluctuations while mitigating extreme spikes that commonly occur in high-demand scenarios. Evaluation using Google Cluster Trace data demonstrates that the proposed model achieves a notable 5.91% reduction in peak energy consumption during the top 10% of workload periods. Visualization of predicted and actual energy traces shows that high-power intervals are significantly smoothed, while the overall workload behavior is preserved, allowing the system to maintain performance without unnecessary energy expenditure. The model's ability to forecast peak loads enables operators to plan resource allocation, reduce periods of extended high-power operation, and avoid potential battery depletion or overheating in edge devices. In addition, rolling average analyses indicate that transient workload bursts are dampened, further contributing to system stability and consistent energy efficiency. These results suggest that Transformer-based workload prediction provides a scalable and reliable approach for energy-efficient resource management in edge computing infrastructures. Overall, this work establishes a foundation for intelligent, energy-conscious design strategies in next-generation edge environments, demonstrating that predictive, AI-driven management can deliver both precise forecasts and measurable reductions in energy-intensive peak periods.