Artificial intelligence for energy-efficient computation offloading in WPT-enabled industrial internet of things

Integrating artificial intelligence (AI) applications in resource-constrained Internet of Things (IoT) systems with intelligent edge computing and Wireless Power Transfer (WPT) is essential for supporting real-time decision-making and sustainable Industrial Internet of Things (IIoT) operations. AI-driven WPT significantly improves the efficiency of time-division multiplexing (TDM) by enabling precise coordination between data offloading and enhancing the overall sustainability and efficiency of the system when combined with intelligent edge computing. To address the stringent battery capacity constraints in WPT systems, a perturbation-based virtual energy queue is proposed to relax the strict energy limitations typically encountered. This mechanism eliminates the need for future system condition prediction, thereby enabling efficient and adaptive real-time scheduling decisions. Furthermore, Dinkelbach’s transformation is employed to reformulate the long-term Energy Efficiency (EE) optimization problem into a tractable drift-plus-penalty framework, effectively reducing latency and ensuring queue stability. To enhance queue stability and intelligent decision-making in online scheduling, this study proposes a hybrid Deep Reinforcement Learning (DRL)–Lyapunov optimization framework that enables adaptive learning by dynamically adjusting the Central Pro-cessing Unit (CPU) frequency to minimize power consumption while satisfying latency constraints derived from the drift-plus-penalty bound. The hybrid DRL–Lyapunov achieves sustainability in the Industrial Internet of Things (IIoT) operation by integrating actor–critic to support the agent learning to obtain an optimal policy in high-dimensional state spaces. Simulation results demonstrate that the proposed hybrid DRL–Lyapunov framework enhances EE by 15–20% compared to a fixed-power baseline, maintaining device battery levels within the optimal range of 55–60%. This approach effectively ensures energy balance, queue stability, and reduced variations in battery dynamics.