⚡ electrical engineering

ENO-Chain: Energy-Neutral Scheduling of Blockchain Participation for Energy-Harvesting IoT Nodes in Smart Grids

This paper introduces ENO-Chain, a Lyapunov-based scheduling policy that dynamically adjusts blockchain participation intensity for energy-harvesting IoT nodes in smart grids to guarantee energy-neutral operation while significantly outperforming static baselines in utility and reliability.

Original authors: Nawaf Abdualaziz Almolhis, Ashish Dixit, Subuhi Kashif Ansari, Huda Fatima, Shamimul Qamar, Salahaldeen Duraibi, Ashwag Alasmari, Salman Arafath Mohammed

Published 2026-10-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nawaf Abdualaziz Almolhis, Ashish Dixit, Subuhi Kashif Ansari, Huda Fatima, Shamimul Qamar, Salahaldeen Duraibi, Ashwag Alasmari, Salman Arafath Mohammed

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a world where the devices that keep our electrical grids running—smart meters on every house, sensors on every solar panel—do not need wires or batteries that require replacement. Instead, they run on the energy they harvest from their surroundings: sunlight, wind, or even the vibrations of the machinery they monitor. This is the promise of the "energy-neutral" smart grid, a system where a device's long-term energy use never exceeds what it collects from the environment. If it does, the device runs out of power and goes silent; if it collects too much, it wastes the surplus. The challenge is that these devices are now being asked to do more than just measure electricity. They are being asked to participate in a digital ledger, a blockchain, to create tamper-proof records of energy use and enable peer-to-peer trading. This new job is computationally heavy and consumes significant power, creating a dangerous conflict: how can a device that must carefully ration its own survival energy also perform a demanding, security-critical task without running out of juice?

Researchers have long known how to manage energy for simple tasks like sensing or sending a signal, but they have struggled to apply those rules to the complex, unpredictable demands of blockchain participation. Previous attempts often treated the energy needed for these digital tasks as a fixed cost or a static rule, failing to account for the fact that the sun might shine brightly one day and be hidden by clouds the next. Without a way to dynamically adjust how much of this digital work a device attempts based on its current energy reserves, the system risks either wasting energy or, worse, causing the device to die and drop offline, breaking the security of the entire network.

To solve this, a team of researchers developed a new control system called ENO-Chain. They approached the problem by treating the intensity of a device's participation in the blockchain not as a fixed setting, but as a flexible dial that could be turned up or down every few minutes. The core of their solution is a two-part strategy. The first part is a strict, mathematical safety guard that ensures the device never spends more energy than it has collected over time. This guard operates like a financial budget that never allows a deficit to grow too large, guaranteeing that the device stays alive regardless of how unpredictable the weather becomes. The second part is a learning system that tries to be clever, using short-term weather forecasts to decide if it can safely turn the dial up to get more value from the blockchain before the sun goes down.

The researchers tested this system in a detailed computer simulation that mimicked a year of changing weather, from bright sunny days to cloudy, low-energy periods. They compared their new system against older methods, including one that simply tried to do the maximum amount of work all the time and another that used a fixed, hand-tuned schedule. The results revealed a surprising truth about how to build these systems. The strict safety guard alone was remarkably effective. It kept the devices running with zero failures, even during the harshest weather, while still achieving significantly better performance than the older, static methods. In fact, the safety guard alone outperformed the hand-tuned baseline by about 21 percent in terms of useful work completed.

However, the addition of the "learning" part of the system did not always help. In many cases, the learning layer, which tried to predict the future to squeeze out extra performance, actually made things slightly worse. It introduced a small but real risk of the device running out of power, and it sometimes reduced the total amount of useful work done compared to the safety guard acting alone. The researchers found that this learning layer only worked well when it was trained on a wide variety of weather conditions. When it was trained only on mixed weather and then tested on very cloudy or very cold days, it failed to adapt correctly, proving that its intelligence was fragile and dependent on its training data.

The study also looked at how the system would behave if thousands of devices were connected at once. They found that the new method was very efficient at managing competition for limited slots in the blockchain network. While older methods wasted many attempts to join the network, the new system learned to back off when it was too crowded or too expensive, saving energy for when it was truly needed. This efficiency held steady even as the number of devices grew from a handful to hundreds.

Ultimately, the paper concludes that the most reliable way to run these energy-harvesting devices is to rely on the strict, mathematically proven safety guard. The learning component can be added to improve performance, but only if it is carefully tested to ensure it does not break the safety guarantee. The researchers showed that in the real world, where weather is unpredictable and energy is scarce, a system that guarantees it will never run out of power is more valuable than one that tries to be clever and occasionally fails. The study serves as a cautionary tale for the future of smart grids: in a world of limited resources, the most sophisticated algorithm is not always the best one; sometimes, the most reliable path is the one that simply knows when to stop.

Technical Summary: ENO-Chain

Problem Statement
Smart grid IoT nodes, such as advanced metering infrastructure (AMI) and distributed energy resource (DER) controllers, are increasingly transitioning to ambient energy harvesting (e.g., solar, vibration) to ensure sustainability and reduce maintenance. This shift imposes the Energy-Neutral Operation (ENO) constraint: a device's long-run energy consumption must not exceed its long-run harvested income to avoid battery depletion. Concurrently, these nodes are required to participate in blockchain ledgers for tamper-evident metering and peer-to-peer settlement.

Existing literature treats these two requirements in isolation. Energy-neutral scheduling research optimizes sensing or communication duty cycles but ignores security-relevant workloads. Conversely, energy-aware blockchain consensus research focuses on reducing compute costs or selecting validators based on static energy snapshots, often assuming mains power or failing to model the stochastic nature of harvesting over time. Consequently, there is a lack of a scheduling framework that jointly manages a stochastic energy budget and a blockchain participation workload while providing a formal guarantee against battery depletion.

Methodology: ENO-Chain
The paper proposes ENO-Chain, a two-layer control policy designed to maximize blockchain participation utility while strictly adhering to ENO constraints. The system models the node's battery state evolution and treats blockchain participation intensity (ranging from sleeping to full validation) as a discrete control variable.

Layer 1: Lyapunov Drift-Plus-Penalty Control
The core of the controller is a Lyapunov optimization policy that does not require prior knowledge of the energy harvesting distribution. It maintains a virtual energy-deficit queue $Q_i(t)$ and minimizes a drift-plus-penalty bound at every time slot. This layer:
- Provides a closed-form $O(1/V)$ – $O(V)$ trade-off between utility maximization and battery deviation.
- Offers a formal guarantee of bounded battery deviation and mean-rate stability, ensuring the time-averaged energy-neutral constraint holds without needing to predict future energy income. However, the paper notes that achieving zero violation rates in practice is contingent on the specific calibration of the trade-off parameter $V$ .
- Operates with low computational complexity, suitable for constrained microcontrollers.
Layer 2: Constrained Actor-Critic Fine-Tuning
A lightweight reinforcement learning (actor-critic) layer refines the decision made by Layer 1. It utilizes short-horizon energy forecasts and recent load patterns to select higher-utility actions. Crucially, this layer is shielded: it is restricted to a "safe set" of actions certified by Layer 1 for the current slot. This design ensures that even if the learned policy fails to converge or acts suboptimally, it cannot violate the safety constraints established by Layer 1 at the moment of selection. However, the paper clarifies that this per-slot certification does not guarantee that a sequence of such actions preserves the long-run violation rate of Layer 1 alone, particularly at higher values of $V$ .

Key Contributions

Formulation: The paper formulates blockchain participation (signing, relaying, validating) as a controllable variable within the canonical ENO battery-drift model, extending the framework beyond traditional sensing workloads.
Controller Design: It introduces a two-layer architecture combining a formal, distribution-free Lyapunov controller with a learned refinement layer, utilizing a shielding pattern adapted from safe reinforcement learning literature.
Rigorous Evaluation Protocol: The authors implement a strict evaluation protocol where the trade-off parameter $V$ is calibrated on a trace disjoint from all evaluation seeds. They also reimplement the closest prior baseline (Saleh and Cevik [11]) at the population level it was originally intended for, rather than as a single-node proxy.
Empirical Findings on Learned Layers: The study identifies and traces a specific failure mode where the learned refinement layer, when trained on a single weather regime, can underperform the safety layer and introduce non-zero violation rates when evaluated on different regimes. This highlights the sensitivity of learned components to training distribution shifts and the risk of eroding safety guarantees at certain operating points.

Results
Evaluated across eight independent traces and four weather regimes (sunny, cloudy, mixed, winter):

Safety and Utility: At the independently calibrated parameter $V^*=10,240$ , the Lyapunov-only layer (Layer 1) achieves a mean utility 21% higher than a hand-tuned static-threshold baseline while maintaining zero violation rates (no battery depletion events) across all eight seeds.
The Learned Layer Trade-off: At this specific operating point ( $V^*$ ), the full two-layer controller (ENO-Chain) exhibited a lower utility and a non-zero violation rate compared to Layer 1 alone. This reversal occurred because the learned layer, trained on a single regime, introduced instability when applied to others, trading safety for utility. The paper notes that at higher $V$ values (e.g., 20,480), the learned layer's refinement can degrade performance even further, despite the formal guarantee of the base layer.
Training Distribution Effect: When the learned layer was retrained on a mixture of all four weather regimes, it successfully matched or exceeded Layer 1's performance in most regimes, confirming that the initial underperformance was a training artifact rather than a fundamental architectural limitation.
Scalability: The method demonstrated stability in per-node utility and reduced validator-slot denial rates as network size scaled from 5 to 200 nodes, outperforming energy-oblivious baselines that suffered high contention waste.

Significance and Claims
The paper modestly claims that its primary contribution is not merely a new controller, but a demonstration of how evaluation protocols can silently favor incorrect ablations. The study shows that calibrating safety-critical hyperparameters on the same data used for testing can mask the fact that a learned refinement layer degrades the formal safety guarantees of the base controller.

The central empirical finding is the asymmetry between the two layers: the Lyapunov layer maintains its formal guarantee (bounded deviation and mean-rate stability) regardless of the learned layer's performance in isolation, but the learned layer's contribution is conditional on its training distribution matching deployment conditions and the specific value of the trade-off parameter $V$ . At certain operating points (specifically high $V$ ), the learned layer can erode the violation rate established by the safety layer. The paper concludes that for deployment, the Lyapunov layer should be relied upon for the neutrality guarantee, and the learned layer should only be added after rigorous validation on held-out data spanning the expected operational conditions. The work emphasizes that in safety-critical, energy-constrained environments, the stability of the formal guarantee is more significant than average performance improvements that may come at the cost of safety.

Technical Summary: ENO-Chain

More like this