Mobile networks save energy by putting base stations to sleep when traffic is low. Papers on this report double-digit savings. One controller based on adaptive dynamic programming (a learned approximation to the optimal sleep/wake policy) claims a 20.1% power reduction at a 98% quality-of-service ratio. That result came from a proprietary simulator fed with undisclosed traffic. Nobody outside the authors can check it.
The reproducibility literature shows that even fully public data does not prevent this kind of fragility. When three top pathology algorithms were rebuilt from their paper descriptions alone, accuracy collapsed because preprocessing details had been omitted. This note asks the same question of green networking, and asks it quantitatively. On open cellular traffic traces, how much of a reported energy saving is decided by the sleep algorithm, and how much by undocumented data-handling choices? Those choices include the time resolution of the traffic, how map grid cells are grouped into "base stations," how train and test periods are split, and whether waking a sleeping radio is assumed to cost anything.
The proposal builds an open, fully specified evaluation pipeline for base-station sleep control. It runs several controllers across a grid of plausible pipeline choices and attributes the variance in measured savings to each factor. The output is not another checklist written by analogy. It is an empirically derived list of the reporting fields that actually move energy results, together with a test of whether gains in traffic-forecast accuracy translate into joules saved at all.
Core Claim
Our hypothesis, testable and falsifiable, is that on open traces, pipeline choices explain at least as much variance in measured savings and quality of service as the choice of controller. We further expect that some of these choices can reverse the ranking between a simple threshold rule and a learned controller. The two prime suspects are resampling 10-minute traffic to hourly and ignoring wake-up energy and latency.
Hourly averaging smooths away the sub-hour surges that cause quality-of-service failures and ping-pong switching, where a cell flips repeatedly between sleep and active. It should therefore inflate both savings and service quality for aggressive policies. Random train/test splits leak temporally adjacent traffic into training. They should flatter forecast-driven controllers.
If the attribution instead shows the algorithm dominating, that is also a useful result. It would mean existing claims are more portable than the transparency critique suggests.
Where This Sits in Prior Work
The motivating evidence comes from three places. Adaptive Dynamic Programming for Energy-Efficient Base Station Cell Switching frames carrier switching as a Markov decision process. It reports 20.1% savings and 98% quality of service, against threshold rules whose quality of service fell "sometimes as low as 68%." All of this was measured on closed data, so the claim cannot be audited.
Reproducibility of deep learning in digital pathology whole slide image analysis reimplemented Camelyon algorithms from text alone. One slide-level accuracy dropped from 92% to 54.7%, and one lesion-detection score dropped from 0.981 to 0.499. The authors traced these drops to unstated preprocessing and proposed a 12-point checklist. We borrow their logic, but there are two differences. They documented failures qualitatively in one reproduction per algorithm, whereas we vary pipeline factors systematically and measure each factor's share of the outcome. And our domain is control-for-energy rather than classification.
Dual Attention-Based Federated Learning for Wireless Traffic Prediction (FedDA) supplies both the data and a forecaster. It uses the open Telecom Italia Milan and Trento call-detail records, resamples them to hourly, and reports MSE gains over FedAtt of up to 10% (Milan) and 48.6% (Trento Internet). It never connects those gains to energy, which is exactly the coupling we test. Its public code, chuanting/FedDA, lets us reuse the forecaster unchanged.
The power side comes from A Parametric Power Model of Upper Mid-Band (FR3) Base Stations for 6G. That paper splits base-station power into digital processing, analog conversion, and amplifiers, with micro-sleep states. Its key finding is that processing, not radio transmission, dominates at high antenna counts. We use it as the load-to-power map.
Two edge-AI measurement papers constrain how much we may trust such models. Unveiling Energy Efficiency in Deep Learning and DeepEn2023 show that operation-count estimates of energy reach only 31.3% accuracy, against 86.2% for a measurement-trained predictor. The main reason is transients such as a roughly 10 ms frequency ramp-up. Their lesson is that the parametric model's instantaneous sleep transitions are an assumption, not a fact.
We cannot measure base-station hardware. So wake-up penalty becomes a swept, explicitly reported factor rather than a hidden constant. This is an honest limitation: our energy numbers are model-relative. Our claims are about sensitivity and ranking, not absolute watts.
The stochastic-geometry work, Optimal area power efficiency in cellular networks and Energy-Efficient Design of MIMO Heterogeneous Networks with Wireless Backhaul, assumes spatially random base stations and users. We use it as a contrast condition. Running the same controllers on synthetic random-placement traffic shows whether real spatial clustering changes conclusions.
Research Program
Pipeline. The pipeline turns Telecom Italia Milan and Trento traffic into per-"base station" load series.
Simulator. An open simulator applies sleep decisions. It maps load to power through the FR3 parametric model and counts quality-of-service violations as unserved demand above remaining capacity after neighbors absorb offloaded traffic.
Pipeline factors. Each step that the literature performs silently becomes an explicit factor:
- temporal resolution: native 10-minute versus hourly;
- spatial aggregation: grid cells grouped into virtual base stations by fixed tiling versus traffic-similarity clustering of the kind FedDA uses;
- missing-cell handling: dropping versus interpolating;
- load mapping: Internet activity alone versus a weighted sum of SMS, call, and Internet;
- evaluation split: random versus contiguous time blocks versus held-out spatial regions;
- wake-up penalty: zero, a short energy-and-latency cost, and a long one;
- switching granularity: carrier versus whole-site.
Factorial design and attribution. We cross the factors in a fractional factorial design and run every controller in every cell. We then attribute variance in energy saving and quality-of-service ratio to each factor and to its interaction with the controller. The attribution uses a standard ANOVA, cross-checked with Shapley-style attribution over factors so that interactions are not misassigned.
Deliverables. We will release the pipeline, the per-cell results, and a short "energy evaluation card." The card lists only those fields whose attributed share exceeds a pre-set threshold. This makes it a data-derived reporting standard rather than a generic one.
Baselines
Do nothing. The naive reference is an always-on network. Every saving is measured against it under the same pipeline cell.
Threshold rule. The simplest serious competitor is the operator-style rule: sleep a carrier when its load falls below a fixed fraction for a fixed dwell time. Its threshold is tuned on the training split with the same budget as the learned controllers. A reviewer will expect this rule to be adequate. The evidence that it fails, the 68% quality-of-service figure, comes only from closed data.
Learned dynamic-programming controller. The strongest existing method is the controller of Adaptive Dynamic Programming for Energy-Efficient Base Station Cell Switching. It combines MLP predictors of power and quality of service with an LSTM handover predictor. Its reported result is a 20.1% power reduction at a 98% quality-of-service ratio, on a proprietary simulator. There is no reported result on Lacuna for this or any cell-switching controller on Telecom Italia data. We reimplement it from the paper description and give it the same training data and hyperparameter search budget as every other method. A reimplementation that falls short of 20.1% is itself a data point. It is not evidence against the original, since the scenarios differ.
Forecast-driven controller. The third learned competitor feeds FedDA forecasts into the tuned threshold rule. This isolates whether better forecasting helps energy.
Oracle. An oracle with perfect future traffic bounds achievable savings.
Key ablation. The ablation that isolates the proposal's main ingredient keeps the controllers fixed and switches only between a single "typical literature" pipeline (hourly, random split, zero wake-up cost) and the full factorial.
"Beating" the baselines is not the goal here. The proposal succeeds if the attribution identifies pipeline factors that change savings by more than the gap between the threshold rule and the learned controller. It succeeds more strongly if they flip which controller wins.
Why not just…?
Why not just release code and data and let others reproduce? Release makes results checkable, but it does not tell readers which of dozens of choices matter. The pathology study shows that open data alone failed to protect results: Camelyon was public, and reproductions still collapsed. Our attribution tells authors which few fields they must report and reviewers which ones to question. If every factor proves negligible, this alternative wins, and we will say so.
Why not just fix one standard preprocessing and move on? A single standard hides the question of whether conclusions survive reasonable variation. If controller rankings are stable across the factorial, a standard is enough and our card reduces to it. If rankings flip, any single standard silently picks a winner. The factorial result decides this directly.
Why not just trust the analytical models, since stochastic geometry already gives optimal densities and sleep thresholds? Those models assume spatially random nodes and instantaneous state changes. The edge-AI measurements show that operation-count estimates of energy miss transients badly, reaching 31.3% accuracy. We test the threat directly by running the controllers on matched synthetic random-placement traffic and comparing their rankings and savings with the real-trace results. If the two agree, the analytical shortcut is vindicated for this purpose.
Why not just sweep the obvious suspect, hourly resampling, and skip the rest? This is a real threat, because one factor might carry nearly all the effect. The factorial is cheap relative to training. Its attribution will reveal whether one factor dominates, and in that case the recommended reporting card shrinks accordingly.
Proposed Experiments and Decisive Checks
Main experiment. We run all controllers over the factorial on Milan, and on Trento as a second city. We report energy saving, quality-of-service ratio, switching count as a proxy for ping-pong, and bits per joule.
Decisive check 1. Does the share of variance from pipeline factors exceed the controller's share? We pre-register a threshold of one half of total explained variance.
Decisive check 2. Do the threshold rule and the learned controller swap order between the "typical literature" pipeline and the 10-minute, time-blocked, nonzero-wake-cost pipeline? We pre-register the directional predictions: hourly resampling and zero wake cost raise apparent savings, and random splits raise apparent quality of service for forecast-driven control.
Forecast-to-energy test. We train FedDA variants and simpler forecasters spanning a range of MSE. We then measure whether forecast MSE rank-correlates with energy saved at fixed quality of service. A weak correlation would show that the MSE gains reported on these traces are not an energy result.
Generalization check. We train on Milan and test on Trento under spatial-block splits. This checks whether learned controllers' advantages survive a city shift, which is the drift problem static documentation cannot capture.
References
- Adaptive Dynamic Programming for Energy-Efficient Base Station Cell Switching
- Reproducibility of deep learning in digital pathology whole slide image analysis
- Dual Attention-Based Federated Learning for Wireless Traffic Prediction
- A Parametric Power Model of Upper Mid-Band (FR3) Base Stations for 6G
- Unveiling Energy Efficiency in Deep Learning: Measurement, Prediction, and Scoring Across Edge Devices
- DeepEn2023: Energy Datasets for Edge Artificial Intelligence
- Optimal area power efficiency in cellular networks
- Energy-Efficient Design of MIMO Heterogeneous Networks With Wireless Backhaul