Conditional What-If Generation for Radio Access Network Energy under Rare and Unseen Contexts

Kevin Wardakhan, Marie Line Alberi-Morel, Jérémie Leguay, Afef Feki, Adrien Sardi, Sébastien Archeveque, Claudiu MihailescuarXiv 2026 · arXiv

TL;DR

Researchers have developed a new generative AI framework to simulate how 5G base stations consume power under hypothetical "what-if" scenarios. By decomposing power consumption into a global prediction and a tunable residual, their model can accurately generalize to rare or entirely unseen network configurations using as little as a single real-world observation.

What Problem Does This Paper Address?

Optimizing the energy efficiency of Radio Access Networks (RANs) is a high-stakes balancing act. Operators want to use features like "carrier shutdown" or "MIMO muting" to save power, but aggressive settings can tank network performance (KPIs). To find the sweet spot, they need "what-if" analysis tools—digital twins that can predict power consumption for various traffic loads and hardware setups before they are deployed.

The catch? Data is sparse. Real-world networks have a "long tail" of configurations. While some base stations are well-represented in training data, many others are "rarely seen" or "unseen." Standard machine learning models often fail in these edge cases because they struggle to estimate the absolute power level of a new site, even if they can guess its temporal patterns.

Methods

The authors propose two novel approaches to bridge this gap:

  1. CN-TimeWeaver (Context-Normalized TimeWeaver): This is a direct conditional generator. It builds on the TimeWeaver diffusion model but adds a "context normalizer." This component predicts site-specific statistics (mean, standard deviation, min/max) directly from hardware metadata to normalize the data before it hits the diffusion denoiser.
  2. C-TimeWeaver (Corrected TimeWeaver): Recognizing that direct generation struggles with unseen sites, the authors introduced a prediction-based framework. It decomposes power consumption into three parts:
  3. Global Prediction: A standard neural network estimates the baseline power.
  4. Residual Correction: This adjusts the "level" (offset) and "scale" (variability). Crucially, the level can be corrected at inference time using a few real-world observations ($K$-shot adaptation) without retraining the model.
  5. Temporal Residual Generation: A generative model (like diffusion or an AR process) creates the "wiggle" or stochastic noise around the predicted baseline.

The architecture for the context-normalized approach is shown below:

The CN-TimeWeaver architecture combines static hardware metadata and dynamic operating context to generate power trajectories.

Datasets

The study utilized two complementary 5G datasets:

  • Private Commercial 5G RAN Dataset: Contains 42,031 samples from 9 base stations (65 distinct hardware groups) with high-resolution 15-minute sampling. This was used to evaluate long-term trajectory generation.
  • Public ITU 5G Dataset: A much larger dataset with over 92,000 hourly measurements from 923 base stations. This provided the diversity needed to test how well the models generalize across hundreds of different sites.

Key Insights and Ideas

  • The "Level" Problem: The biggest hurdle for generative models in unseen contexts isn't the shape of the power curve—it's the vertical offset. As shown in Figure 4, the error in predicting the "mean" power level spikes significantly for unseen groups.
  • Metadata vs. Observations: Hardware metadata (like the 5G band or number of cells) is great for predicting the scale of power fluctuations, but it's surprisingly poor at predicting the absolute level.
  • The Power of One: One of the most striking findings is that just one target observation ($K=1$) at inference time can "anchor" the model, shifting the entire generated trajectory to the correct level and cutting error rates in half.
Figure 5 illustrates how a single observation (K=1) corrects the vertical shift of the generated trajectory for a challenging unseen group.

Key Results

  • Well-Seen Contexts: Both proposed models significantly outperformed the baseline (TimeWeaver). CN-TimeWeaver achieved the best temporal similarity (DTW).
  • Unseen Contexts: The prediction-based C-TimeWeaver was the clear winner. With just one observation ($K=1$), it reduced the median Mean Absolute Error (MAE) by 50% compared to the state-of-the-art.
  • Unseen Operating Regimes: C-TimeWeaver also proved robust when asked to generate data for high-traffic scenarios that were entirely missing from that specific site's training history.

The trade-offs between these methods across different training exposures are summarized in the authors' evaluation:

Fig. 6. Median MAE across training-exposure regimes and inference support. Few-shot results (K > 0) are reported for unseen groups. C-AR(3) is shown only in the unseen setting.

Limitations and Future Work

While the diffusion-based models are highly accurate, they are computationally expensive, taking about 16 seconds to generate a trajectory compared to just 5 milliseconds for simpler autoregressive (AR) models.

The authors suggest that future work should focus on:

  • Improving "zero-shot" correction (predicting the level accurately without any observations).
  • Better probabilistic calibration.
  • Applying these models to more complex "what-if" scenarios involving multi-layered energy-saving policies.

Summary updated

Abstract

Energy-saving mechanisms in RANs must balance energy reduction and network performance under varying traffic conditions. Conditional generative models support what-if analysis by generating plausible power-consumption trajectories for radio configurations and operating scenarios, helping operators evaluate energy-saving policies before deployment. However, representative measurements are costly and operationally constrained, so historical data cover only limited configurations and conditions, leaving some contexts poorly represented or unseen during training. In this paper, we study conditional generation of RAN power-consumption trajectories under decreasing training coverage, from well-represented to rare and unseen configurations. We first introduce a context-normalized conditional generator combining temporally aligned conditioning with configuration-dependent target normalization. It improves generation for well-represented configurations but becomes less reliable as training exposure decreases because target consumption is harder to estimate. We therefore introduce a prediction-based generative framework that decomposes power consumption into global prediction, residual level and scale correction, and temporal residual generation. Experiments on a commercial 5G RAN and a public dataset show that hardware metadata provides transferable information about residual scale, while limited target observations at inference improve generation through residual-level correction without retraining. An ablation study confirms that residual-level correction provides most of the gain. Both approaches outperform state-of-the-art methods for well-represented configurations, while the prediction-based approach also generalizes to rare and unseen contexts and, with one target observation, reduces median MAE by about 50% compared with TimeWeaver.
BibTeX
@misc{wardakhan2026conditional,
  author = {Kevin Wardakhan and Marie Line Alberi-Morel and Jérémie Leguay and Afef Feki and Adrien Sardi and Sébastien Archeveque and Claudiu Mihailescu},
  title = {{Conditional What-If Generation for Radio Access Network Energy under Rare and Unseen Contexts}},
  year = {2026},
  eprint = {2611.24208},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.24208},
}
Figure