Energiebewustzijn voor het Internet of Things

Abstract

The Internet of Things (IoT) is a pervasive ecosystem, currently comprising more than 20 billion connected devices. This widespread deployment of devices allows ubiquitous sensing, communication, and new levels of human-computer interaction. However, these devices must contend with unpredictable user demands and highly variable power supply and demand, making reliable long-term operation difficult to guarantee. The traditional response of masking that unpredictability with oversized, finite batteries incurs significant maintenance costs and generates toxic electronic waste, driving the need for sustainable, energy-aware IoT operation. Even where energy supply is stable, interactive deployments face sporadic, user-driven demands that conventional, schedule-based power management cannot anticipate. Establishing the real-time energy awareness required to react to such dynamic demands is itself costly, as the dedicated monitoring hardware and software needed for visibility draw continuously from the very reserves they seek to protect. Transitioning to battery-free, ambient energy harvesting addresses the environmental concerns but only deepens these challenges: an unpredictable supply must now dictate software execution to avoid failures. Contemporary adaptive schedulers and energy models, meanwhile, typically depend on static lab benchmarks, prolonged off-device training cycles, or hand-tuned parameters, leading to poor performance or high labour costs for real-world deployments. Against the backdrop of these challenges, this dissertation contributes four building blocks towards sustainable, self-optimising operation on energy-constrained embedded systems. The first introduces the Citizen Dialog Kit (CDK), an integrated hardware and software platform for autonomous in-situ public surveying. By pairing low-power cellular networking with an E-Ink display and a compact survey bytecode interpreted on-device, CDK absorbs unpredictable interaction loads, dramatically reduces network bandwidth requirements, and achieves multi-year battery lifetimes in realistic deployments. The second contribution, miniMaP, delivers the high-fidelity in-situ measurement and profiling needed to validate such lifetime claims without itself dominating the energy budget. By dynamically aligning the power states of a monitoring integrated circuit with software execution, miniMaP attributes per-context statistics, including mean, variance, and peak power, back to the application threads that produced them, at sampling rates of up to 20kHz. At more moderate sampling rates, CPU active-time overhead falls as low as 12.15%, with miniMaP's own power draw down to 2.95mW. The third contribution, the diurnally optimised AsTAR++ scheduling algorithm developed within the Flute battery-free platform, transitions off-the-shelf IoT hardware to sustainable operation under highly variable ambient power. By matching software execution to available environmental energy and exploiting day-night solar patterns to safely draw down supercapacitor charge overnight, AsTAR++ sustains continuous operation across week-long indoor and outdoor solar deployments without prior energy benchmarking, lifting task throughput to as much as 17 times the AsTAR baseline while preserving a 10% safety margin above brownout. The fourth contribution, the Model-Augmented Q-Learning Scheduler (MAQS), removes the manual parameter tuning or off-device training that previous schedulers require. By formulating energy management as a Markov Decision Process and introducing Phantom Q-updates, which propagate reward signals across multiple virtual actions from each experience, MAQS allows devices to learn optimal scheduling policies entirely in-situ within a memory footprint of less than half a kilobyte. Evaluation against real-world solar irradiance datasets shows that MAQS achieves a superior balance of reliability and throughput compared to AsTAR and AsTAR++, eliminating power failures after a rapid convergence period while sustaining higher task rates than either baseline. Together, these contributions embed in-situ energy awareness across the hardware-software stack: bounding interactive energy costs, attributing fine-grained consumption back to software contexts, sustaining off-the-shelf nodes on harvested supply, and replacing hand-tuned schedulers with on-device learning. They demonstrate that the longstanding reliance on ex-situ modelling and manual parameter tuning is no longer a necessary compromise; resource-constrained devices can instead be granted the means to measure, reason over, and adapt to their own local conditions. In doing so, this dissertation lays a foundation for sustainable, maintenance-free pervasive computing at scale, built not on denser finite batteries or heavyweight predictive models but on lightweight, reactive intelligence at the edge.