Unified Energy Efficiency Optimization Under Uncertainty in EH-WSNs

Abstract

Enhancing sensor longevity is crucial for the effective operation of wireless sensor networks (WSNs). Energy harvesting (EH) sensors address this challenge by harvesting ambient energy and extending operational lifespans. However, without energy-efficient resource allocation, the dynamic nature of EH rates can disrupt node operations and degrade network performance. Existing studies have separately addressed crucial challenges in resource allocation scenarios under uncertainties in EH-WSNs. In contrast, our work presents a unified framework that optimizes energy efficiency (EE) under uncertainty by adopting an intelligent probabilistic approach that integrates energy-efficient resource allocation with EH dynamics for a unified solution. We achieve this by reformulating the conventional deterministic optimization problem (DOP) into a set of probabilistic optimization problems (POPs), encompassing stochastic, robust, and chance-constrained models. To address the complexity of solving non-convex POPs in uncertain and dynamic environments, we propose a custom-tailored sample average approximation (SAA)-assisted deep reinforcement learning (DRL) optimizer employing a built-in Deep Q-Network (DQN) agent. Leveraging the inherent adaptivity of SAA-assisted DRL, the proposed framework dynamically adjusts to varying environmental conditions, enabling unified and efficient optimization in the face of uncertainty. Furthermore, to ensure a robust and credible benchmark, we also employ Double DQN (DDQN) as a DRL baseline, enabling evaluation of our method against multiple variants and facilitating a clear comparison of convergence behaviors. Simulation results demonstrate that our unified probabilistic framework achieves near-optimal performance in terms of mean absolute error and convergence rate, even in the presence of EH uncertainties.