Abstract
Abstract
Agricultural Wireless Sensor Networks (WSNs) are increasingly required to operate over long periods under dynamic environmental conditions while relying on strictly constrained energy resources. Although hierarchical clustering and UAV-assisted mobile sinks can reduce communication overhead, the operational lifetime of battery-powered networks remains fundamentally limited. This paper proposes EH-SWADS, an energy-harvesting extension of our previously proposed weather-aware UAV-assisted WSN architecture, SWADS, designed for sustainable, long-term agricultural monitoring. EH-SWADS integrates three core components: (i) an LSTM-based weather prediction module that governs proactive handover between a mobile UAV sink and a fixed ground sink during adverse weather, (ii) a reinforcement learning-based cluster-head selection mechanism enhanced with energy-harvesting awareness, and (iii) solar-powered sensor nodes capable of replenishing energy from ambient irradiance. Unlike conventional approaches that treat harvested energy as a passive buffer, EH-SWADS explicitly incorporates the harvested-to-consumed energy balance into the learning process. Extensive MATLAB simulations across 20,000 rounds demonstrate that solar energy harvesting is the dominant driver of the observed lifetime extension relative to non-harvesting baselines, including our previously reported SWADS architecture (approximately 668% in first-node death relative to a matched non-harvesting baseline, both evaluated under a real, time-varying weather trace, using a corrected reward formulation described below; overall cumulative throughput improved by approximately 1.93×), while a reinforcement-learning-based cluster-head selection mechanism that explicitly rewards harvest-awareness and rotation fairness performs statistically comparably to an otherwise-identical EH-unaware reward weighting (within approximately 5% either direction across seeds) once a harvest-rate normalization flaw and an energy-blind fairness term are corrected; the initial, uncorrected formulation underperformed the EH-unaware weighting by 24–58%, underscoring the importance of validating reward-shaping terms under realistic, time-varying weather rather than idealized conditions.