💻 computer science

Leakage-Resistant and Uncertainty-Aware Machine Learning for Carrier-Grade Router Power Prediction Using Real Traffic Telemetry

This study develops leakage-resistant, uncertainty-aware machine learning models using real traffic telemetry to accurately predict carrier-grade router power consumption, achieving high precision with lightweight, router-specific approaches suitable for Network Digital Twin and zero-touch green networking applications.

Original authors: Clinton Amponsah, Kyiewu Bernard, Linda Bessa-Simmons

Published 2026-10-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Clinton Amponsah, Kyiewu Bernard, Linda Bessa-Simmons

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

The internet is often imagined as a cloud of invisible data, but the reality is a vast, humming network of physical machines that consume enormous amounts of electricity. Every time a video streams, a message sends, or a file downloads, it passes through specialized computers called routers, which act as the traffic directors of the digital world. These devices must stay awake and active to keep the flow moving, and as global data traffic grows, the energy required to power them has become a critical concern for network operators and the environment alike. To reduce this energy use without slowing down the internet, engineers need to know exactly how much power a router uses at any given moment based on how much traffic it is handling. The challenge is that power consumption is not a simple, fixed number; it fluctuates with the type and volume of data passing through, and different router models behave in unique ways. If operators could predict these power needs with high accuracy, they could make smarter decisions about when to switch equipment on or off, or how to route data more efficiently, potentially saving significant energy without compromising service.

A team of researchers set out to solve this puzzle by building a system that could learn the specific power habits of real-world, carrier-grade routers using actual traffic data. Instead of relying on theoretical estimates or simulated numbers, they gathered over 23,000 real measurements from two different families of high-end routers. These routers were subjected to controlled experiments where the researchers gradually increased and then decreased the amount of data flowing through them, mimicking the natural rise and fall of internet traffic throughout a day. The goal was to create a mathematical model that could look at the current traffic conditions—such as how full the router's pipes are and how big the data packets are—and accurately guess the exact wattage the router was consuming at that instant.

The researchers faced a significant hurdle common in machine learning: the danger of the model "overfitting" by memorizing the data rather than learning the underlying rules. Because the measurements were taken every five seconds, the data points were highly connected; a reading taken one second after another was almost identical to the first. If the researchers simply shuffled all the data randomly to train and test their models, the computer might just be recognizing patterns it had already seen, leading to falsely optimistic results. To prevent this, the team designed a strict testing method where they trained the models on the data from the rising traffic phase and then tested them exclusively on the falling traffic phase. This ensured the models had to predict power usage for a future state they had never seen before, providing a much more honest measure of their accuracy.

The results revealed that there is no single formula that works for every router. The two families of routers studied behaved quite differently. One model, a larger and more powerful router, consumed a massive amount of electricity overall, but its power usage barely changed even when the traffic load varied. For this device, the researchers found that a complex, curved mathematical relationship was needed to predict its power use, achieving an average error of just over 5 watts. The second router, a smaller model, used far less electricity in total, but its power consumption rose and fell more noticeably with the traffic. For this machine, a much simpler, straight-line relationship was sufficient to make accurate predictions, with an average error of less than 1.5 watts. This distinction is crucial because it proves that energy-saving strategies cannot be "one size fits all"; they must be tailored to the specific hardware in use.

Beyond just predicting the numbers, the study also focused on how certain the predictions were. In the real world, a prediction is only useful if you know how much you can trust it. The researchers used statistical techniques to create a safety margin around every prediction, essentially drawing a band of uncertainty. For the smaller router, this band was very tight, meaning the prediction was highly reliable. For the larger router, the band was wider, reflecting the fact that its power usage was harder to pin down perfectly. This approach allows network managers to make decisions with eyes wide open; if a proposed energy-saving action relies on a predicted saving that is smaller than the margin of uncertainty, the system knows to hold off, preventing risky or unnecessary changes.

The study confirmed that the most important factor in predicting power use is simply how much of the router's capacity is being used at any given time. While the size of the data packets and the speed at which they arrive did add some helpful detail, they were secondary to the overall traffic load. The researchers also showed that adding too many complex features to the model did not necessarily make it better, suggesting that simple, lightweight models are often the most practical for real-world deployment. These findings provide a solid foundation for "zero-touch" network management, where systems automatically adjust to save energy without human intervention, and for "digital twins," which are virtual copies of physical networks used to test new ideas safely before applying them to the real world.

Ultimately, this work demonstrates that we can understand and predict the energy footprint of complex network equipment with a high degree of precision, provided we respect the unique nature of each device and the flow of time in the data. By moving away from generic assumptions and using real, measured data with rigorous testing, the researchers have shown that accurate power prediction is possible. This clarity is the first step toward a greener internet, where energy is saved not by guessing, but by knowing exactly what is happening inside the machines that keep the world connected.

Technical Summary: Leakage-Resistant and Uncertainty-Aware Machine Learning for Carrier-Grade Router Power Prediction

Problem Statement
Energy-aware network management requires reliable models that translate operational traffic conditions into equipment-level power estimates. However, existing green-networking research often relies on simulated power functions, randomly partitions temporally correlated measurements, or reports point accuracy without quantifying uncertainty. These practices risk overestimating model generalization, which is critical for closed-loop control systems like zero-touch management and Network Digital Twins (NDT). Inaccurate predictions in such environments could trigger unnecessary rerouting, activate redundant equipment, or violate service requirements. Furthermore, energy-efficient routing decisions must balance power reduction against congestion and reliability, necessitating accurate, hardware-specific power functions rather than generic assumptions.

Methodology
The study employed a quantitative experimental design using real traffic telemetry from two carrier-grade router families (RA and RB). The dataset comprised 23,705 observations collected at approximately five-second intervals, covering traffic-load variations, packet-length changes, and multiple concurrent-flow conditions.

Feature Engineering: The model utilized throughput utilization ( $u_t$ ), packet size ( $s_t$ ), and a logarithmically transformed packet rate ( $\log(1 + r_t)$ ). An interaction term ( $u_t \log(1 + r_t)$ ) was constructed to capture the combined influence of traffic load and packet-processing intensity.
Leakage-Resistant Validation: To prevent data leakage inherent in time-series data, the study avoided random row-level splits. Instead, it separated complete traffic-ramp cycles. The "upward" traffic ramp (0% to 100% load) served as the training and model selection set, while the subsequent "downward" ramp (100% to 0% load) formed an independent temporal test set. Model selection within the training set used five-fold grouped cross-validation, ensuring that repeated observations from identical operating states remained within the same fold.
Model Selection: Six regression approaches were compared: a mean-value baseline, linear regression, second-order polynomial ridge regression, Random Forest, Extra Trees, and Histogram Gradient Boosting.
Uncertainty Quantification: The study quantified sampling uncertainty using a moving-block bootstrap (500 repetitions) and predictive uncertainty using split-conformal prediction intervals to generate statistically calibrated 90% prediction intervals.
Ablation Analysis: A post-hoc feature-set ablation evaluated the incremental contribution of packet-rate transformations and interaction terms against a baseline of traffic load and packet size alone.

Key Contributions

Hardware-Specific Modeling: Development of separate, lightweight power-prediction models for two distinct carrier-grade router families, acknowledging that a universal power function is inappropriate due to hardware heterogeneity.
Leakage-Resistant Design: Introduction of a validation framework that separates complete traffic-ramp cycles and groups repeated operating states, preventing the over-optimistic performance estimates common in random splits of temporally correlated data.
Uncertainty-Aware Prediction: Quantification of both sampling and predictive uncertainty using moving-block bootstrap confidence intervals and split-conformal intervals, providing a statistically calibrated basis for decision-making.
Feature Contribution Analysis: Evaluation of the incremental value of packet-rate transformations, supporting parsimonious deployment in Network Digital Twin and zero-touch management environments.

Results

Model Performance:
- Router RA: Second-order polynomial ridge regression performed best, achieving a Mean Absolute Error (MAE) of 5.043 W, Root Mean Squared Error (RMSE) of 7.151 W, and an $R^2$ of 0.645 on the independent downward-ramp test set. The model captured the central load-dependent trend but showed higher errors during packet-length variations.
- Router RB: Linear regression performed best, achieving an MAE of 1.383 W, RMSE of 1.876 W, and an $R^2$ of 0.823. RB exhibited a more linear load-power relationship compared to RA.
Feature Importance: Traffic utilization was confirmed as the dominant predictor. Packet-size and packet-rate transformations provided modest but measurable gains in prediction accuracy, particularly for Router RA.
Uncertainty Calibration: The split-conformal intervals achieved empirical coverage close to the nominal 90% target (88.55% for RA and 90.52% for RB). The bootstrap confidence intervals confirmed the stability of the error estimates under temporal dependence.
Hardware Differences: Router RA operated at a significantly higher power level (~~720 W) with a small dynamic range (~~28 W, ~~4% increase over idle), dominated by a fixed-power component. Router RB operated at a lower level (~~76 W) with a larger proportional response to load (~17–19% increase).

Significance and Claims
The paper claims that carrier-grade router power consumption can be effectively predicted from a compact set of traffic variables when model development rigorously controls for temporal leakage and hardware heterogeneity. The findings support the integration of lightweight, router-specific power models into Network Digital Twins and zero-touch green-networking applications.

Crucially, the authors maintain a modest scope regarding the implications:

The study distinguishes between prediction accuracy and demonstrated network-level energy savings. While the models provide accurate equipment-level estimates, the paper does not claim to have established network-wide energy savings. Such claims would require additional evaluation of routing strategies, interface consolidation, sleep states, and trade-offs involving latency and congestion.
The results suggest that for Network Digital Twins, equipment-specific characteristics must be preserved rather than assuming universal power functions.
The uncertainty-aware decision rule proposed implies that energy-saving actions should not be triggered if the expected savings are smaller than the associated predictive uncertainty interval.

Limitations
The study is limited to two router families and two controlled traffic conditions, restricting broad generalization to other hardware or operational scenarios. The telemetry data excluded routing paths, interface states, temperature, protocol mix, and control-plane activity. Consequently, the models estimate instantaneous equipment power but cannot directly establish end-to-end network energy savings without further integration into a broader control system.

Technical Summary: Leakage-Resistant and Uncertainty-Aware Machine Learning for Carrier-Grade Router Power Prediction

More like this