Reinforcement learning-based scheduling in self-powered IoT systems with piezoelectric harvesting

An energy-aware task scheduling framework based on reinforcement learning is designed to optimize energy consumption, latency, throughput, and operational cost of self-powered Internet of Things (IoT) systems that operate in a cloud-edge setting, and whose energy harvesting is dynamically adjusted. The methodology combines piezoelectric energy harvesting, adaptive task offloading, multi-objective optimization, and intelligent routing to facilitate real-time decision-making. An average harvested energy of 0.82 mJ is obtained using the system with a battery stability rate of 68, which can sustain operation. Task allocation favors edge execution (46.7%), balancing between delay (18.9 ms) and energy (0.36 mJ), and remains low in terms of operational cost at $0.021 per slot. The reinforcement learning model is able to converge to a stable reward of 0.91 in 850 episodes, and this means that the policy learning is efficient. Latency reduction, energy reduction, and throughput enhancement have improvements of 37.6, 37.7, and 37.9%, respectively, in performance evaluation. The system also delivers packets with a 97.2% ratio and prolongs the network lifetime by 47.9%. The architecture is scalable, reliable, and sustainable to support advanced IoT applications.