Adaptive CPU frequency scaling for energy-efficient and sustainable edge computing under renewable energy uncertainty

Key findings from this study

This research indicates that:

• A Double Deep Q-Network agent reduced frequency prediction error by 35% compared to traditional convex optimization methods.
• Dynamic CPU frequency adjustment achieved optimal performance at 1.8 GHz while balancing energy consumption and execution latency under renewable energy variability.
• Reinforcement learning enabled real-time adaptive policies that conventional optimization techniques cannot achieve in stochastic, time-varying environments.

Overview

Edge servers powered by renewable energy sources require dynamic CPU frequency scaling to balance computational demands against intermittent energy availability. Traditional convex optimization methods fail to capture stochastic fluctuations in renewable supply and variable workloads. This research integrates convex optimization with Deep Q-Network reinforcement learning to enable real-time CPU frequency adjustment under energy uncertainty.

Methods and approach

The study formulates CPU frequency control as a stochastic Markov Decision Process and employs a Double Deep Q-Network agent to jointly optimize frequency scaling, latency, and energy efficiency. The DDQN framework learns adaptive policies through real-time environmental interaction, enabling dynamic frequency adjustments constrained by renewable energy availability.

Results

The DDQN policy reduced prediction error by 35% relative to conventional methods and identified an optimal balance point at 1.8 GHz with a lower energy-latency product. Real-time frequency adjustment in response to fluctuating energy availability enhanced energy storage utilization, CPU throughput, and edge resource efficiency. The reinforcement learning approach achieved system stability and performance levels that traditional convex optimization methods cannot attain.

Implications

Integrating reinforcement learning with optimization techniques enables edge computing infrastructure to operate sustainably under renewable energy constraints. The approach directly supports the development of energy-efficient IoT ecosystems by improving dynamic resource allocation without requiring accurate energy predictions. Practical deployment of DDQN-based CPU frequency scaling could reduce operational energy costs and carbon footprints in distributed edge computing environments.

The methodology generalizes beyond renewable-powered edge servers to other dynamic resource allocation problems with stochastic supply and variable demand. The demonstrated 35% improvement in prediction error suggests that learning-based approaches outperform static optimization when system parameters change frequently. These findings advance the technical foundations for sustainable computing infrastructure at the edge of networks.

Scope and limitations

This summary is based on the study abstract and available metadata. It does not include a full analysis of the complete paper, supplementary materials, or underlying datasets unless explicitly stated. Findings should be interpreted in the context of the original publication.