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

The inherent variability and intermittency of Renewable Energy (RE) sources present a major challenge to maintaining efficient and sustainable Internet of Things (IoT) ecosystems powered by Edge Servers (ES) and adaptive Central Processing Units (CPUs). Inaccurate dynamic energy prediction and suboptimal offline CPU frequency adjustments lead to degraded service quality and increased energy costs, despite the need to optimize Sustainable Edge Computing (SEC) execution latency. Traditional convex optimization techniques, such as Newton or Lambert-based methods are effective for static and deterministic CPU frequency tuning but struggle to capture the stochastic, time-varying behavior of RE supply and dynamic workloads. Therefore, enhancing SEC requires a more adaptive mechanism capable of real-time CPU frequency control under uncertain energy availability. To address this limitation, this paper integrates convex optimization with a Double Deep Q-Network (DDQN) framework. The DDQN agent is designed to jointly optimize CPU frequency scaling, computation latency, and energy efficiency by formulating the CPU control problem as a stochastic Markov Decision Process, enabling dynamic frequency adjustment and execution-time optimization under energy constraints. Through real-time interaction with the dynamic environment, the DDQN agent learns near-optimal frequency adjustment policies that adaptively balance computational latency, energy consumption, and RE availability, thereby achieving system stability and performance levels that conventional convex optimization methods cannot attain. Simulation results demonstrate that the proposed DDQN policy reduces the prediction error by 35% and achieves an optimal balance at 1.8 GHz with a lower energy–latency product. By dynamically adjusting CPU frequency in response to fluctuating energy availability, the DDQN agent enhances energy storage utilization, CPU throughput, and edge resource efficiency. © 2026 The Author(s).