The rapid advancement of Tiny Machine Learning (TinyML) is enabling intelligent inference on highly resource- and energy-constrained Internet of Things (IoT) devices. However, sustaining computation over long operational lifetimes remains a major barrier to scalable, low-maintenance deployments, as battery replacement cost and environmental burden often dominate total cost of ownership. To address this challenge, we present a co-designed, hybrid green energy-harvesting TinyML framework that integrates photovoltaic (PV) and ambient radio-frequency (RF) energy sources with supercapacitor buffering and reinforcement-learning (RL) based inference scheduling for indoor air-quality monitoring applications. We first review indoor ambient energy sources and motivate PV + RF hybridisation to mitigate intermittency and improve availability. A dual-input powermanagement architecture with maximum power point tracking (MPPT), impedance-aware conditioning, and supercapacitor-based energy buffering is developed to support energy-neutral operation, with buffer voltage serving as an observable proxy for the node energy state. We then profile three representative TinyML models—Decision Tree, Random Forest, and a Tiny Neural Network— demonstrating the accuracy–energy trade-off that constrains harvesting-powered edge intelligence: 85.7%, 90.9%, and 92.2% accuracy at 5 mJ, 7 mJ, and 9 mJ per inference, respectively. Inference scheduling is formulated as a Markov decision process and solved using tabular Q-learning, which adapts inference execution based on capacitor voltage and historical harvested-power trends to balance predictive utility against brown-out risk. MATLAB simulations over a 12-hour stochastic indoor illumination cycle show that a 25cm2 PV tile under 500 lx lighting, supplemented with ambient RF at − 10dBm, delivers a mean harvested power of approximately 1.2mW. The proposed RL policy maintains the supercapacitor within safe operating limits, achieves near-continuous availability, and attains more than 95% probability of energy-neutral operation, corroborated via Monte Carlo analysis under environmental perturbations. Overall, the results establish a scalable methodology for self-sustaining, battery-less intelligent IoT nodes capable of continuous operation, advancing sustainable and autonomous environmental sensing networks.
Uko et al. (Sat,) studied this question.