EMFusion: Uncertainty-Aware Conditional Diffusion Model for Multivariate Narrow-band Exposure Forecasting

AI summary

Overview

Research area: Machine learning for wireless networks — specifically, generative time-series forecasting of radio-frequency electromagnetic field (EMF) exposure using conditional diffusion models.

Technical level: Advanced. The paper assumes familiarity with denoising diffusion probabilistic models (DDPMs), U-Net architectures, cross-attention, and probabilistic forecasting metrics such as CRPS and NRMSE.

Scope: This paper proposes EMFusion, a conditional diffusion framework for multivariate, narrow-band, uncertainty-aware EMF exposure forecasting, and evaluates it on Italian multi-operator, multi-frequency measurement data.

What This Paper Is About

Most existing EMF forecasting work predicts a single aggregate (wideband) exposure value as a deterministic point estimate, which is insufficient for regulators who must enforce exposure limits band by band and for operators who need to know which frequency, technology, or operator drives exposure. This paper builds a generative model that forecasts EMF levels across many narrow-band frequency channels simultaneously, conditions the forecast on contextual factors such as working hours, working days, and season, and produces probabilistic prediction intervals rather than a single number.

Key Contributions

1. A conditional diffusion model for multivariate EMF forecasting. EMFusion uses a residual U-Net backbone with time-step embeddings, plus a cross-attention mechanism that injects external conditions (time of day, day of the week, holiday/working days, working hours, seasonal trends) directly into both encoder and decoder feature maps rather than simply concatenating them.

2. An imputation-based sampling strategy. Forecasting is framed as a structural inpainting task: during reverse diffusion, observed past values are re-noised and clamped back into the sample using a mask, so the model completes the "missing" future portion of a temporal sequence while remaining temporally coherent even when measurements are irregular or missing.

3. Probabilistic interval estimation via Kernel Density Estimation (KDE). Instead of assuming Gaussian errors, the ensemble of generated trajectories is treated as samples from a conditional distribution, and KDE reconstructs continuous densities at each time step, yielding empirical prediction intervals.

4. Flexibility and empirical validation on narrow-band data. EMFusion can run in either univariate or multivariate mode, and is evaluated on narrow-band EMF datasets spanning multiple channels in the 9 kHz–6 GHz range covering major operators in Italy and cellular technologies.

Main Findings

• Best configuration uses working-hour conditioning: EMFusion with the working-hours contextual information outperforms baseline models both with and without conditioning.

• Reported gains over the best baseline: 23.85% improvement in continuous ranked probability score (CRPS) and 13.93% improvement in normalized root mean square error (NRMSE).

• Conditional beats unconditional, and multivariate beats univariate: The paper reports that conditioning improves results over the unconditional variant, and the multivariate formulation improves over the univariate variant. The paper states that among the exogenous conditions tested, the working-hour condition performed best; whether any condition underperformed is not reported in the provided content.

• Guidance scale set to zero: The authors use classifier-free guidance but found through hyperparameter tuning that a guidance scale of s_c = 0 is best, noting that positive guidance scales tend to over-concentrate the posterior and narrow prediction intervals.

• Frequency coverage of the dataset: The Italian measurement dataset covers 2G, 3G, 4G, and 5G bands across operators labeled Iliad, TIM, VF, and W3, with channel frequencies listed between 763.0 MHz and 3760.0 MHz in Table II, while the paper describes the experimental range more broadly as 9 kHz–6 GHz.

• Dataset size, exact baseline model names, diffusion step count T, and training hyperparameters are not reported in the provided content.

Methodology in Plain English

The model learns to turn random noise into a plausible future EMF trace through many small denoising steps. The full input is a matrix holding both the known past window and the unknown future window across all frequency channels. During training, Gaussian noise is added to this matrix at a random step, and a U-Net is trained to predict the noise that was added, with a mean-squared-error loss.

Two mechanisms guide the process. First, information about the diffusion time step is encoded with sinusoidal embeddings and added into each residual block, so the network behaves differently at different noise levels. Second, the contextual conditions are fed in through cross-attention: the feature map supplies the queries, while the condition vector supplies the keys and values, letting the model focus on whichever context is relevant at each processing stage.

At inference, the process runs in reverse from pure noise. To keep the forecast anchored to reality, the known past values are re-noised at every step and clamped back into the sample via a mask — the inpainting idea. Running this sampling many times produces an ensemble of plausible futures, from which KDE builds prediction intervals without assuming a Gaussian distribution. The authors also use classifier-free guidance with a tuned scale, settling on zero.

Why This Matters

Impact on research. The paper argues that conditional diffusion models had not previously been applied to time-series forecasting in wireless networks generally, nor to EMF forecasting specifically. It reframes EMF forecasting from point prediction of aggregate wideband power to joint probabilistic prediction across frequency channels, operators, and technologies — a shift that also connects forecasting to network-control formulations where predicted exposure acts as a future constraint.

Real-world applications:

• Regulatory compliance: Environmental protection agencies can enforce maximum exposure limits per frequency band, and forward-looking forecasts can turn periodic reporting from retrospective documentation into proactive decision support.
• Network planning and optimization: Operators can anticipate when and in which bands exposure may approach limits, enabling threshold-based alerts, adaptive scheduling, and exposure-aware resource management before violations occur.
• Technology and operator attribution: Narrow-band breakdown lets analysts identify dominant exposure sources and quantify the contribution of specific technologies such as 3G, 4G, and 5G deployments.
• Handling imperfect monitoring data: The imputation-based sampling strategy targets sensor outages and irregular sampling, which the paper identifies as a common real-world constraint.

Industry relevance. Mobile network operators, spectrum regulators, and EMF monitoring bodies all depend on exposure estimates. The paper's framing of predicted exposure as a constraint in utility maximization under QoS and power constraints maps directly onto existing radio resource management pipelines, and the finding that working-hour context is the most informative condition suggests a practical, low-cost feature for deployment.

Future Directions

• Closed-loop network-control validation. The paper explicitly leaves validation in a control loop — using predicted exposure as an operational constraint alongside utility maximization and QoS constraints — to future work.
• Conditioning beyond calendar features. The model supports conditioning on operational metadata and exogenous covariates; other candidates such as population density (used in prior EMF work) or weather (used in other wireless forecasting work) are not tested here.
• Extending the condition set and evaluating more. The paper reports that working hours outperformed among the tested conditions, which raises the question of which other contextual signals could improve accuracy.
• Broader robustness and uncertainty evaluation. The paper positions uncertainty quantification and robustness to incomplete observations as underexplored in prior EMF work; full reporting of calibration quality, extreme-value behavior, and performance across all frequency channels is not detailed in the provided content.

Target Audience

Researchers and graduate students in wireless communications, RF exposure assessment, and generative machine learning; regulators and standards bodies working on EMF compliance; and network engineers or data scientists at telecom operators who build forecasting and resource-management pipelines and need probabilistic, band-specific exposure predictions.

Authors’ abstract

The rapid growth in wireless infrastructure has increased the need to accurately estimate and forecast electromagnetic field (EMF) levels to ensure ongoing compliance, assess potential health impacts, and support efficient network planning. While existing studies rely on univariate forecasting of wideband aggregate EMF data, multivariate narrow-band EMF forecasting is needed to capture the inter-operator and inter-frequency variations essential for proactive network planning. To this end, this paper introduces EMFusion, a conditional diffusion-based EMF forecasting framework that integrates diverse contextual factors, such as time of day, season, and holidays, while providing uncertainty-aware probabilistic forecasts. The proposed architecture features a residual U-Net backbone enhanced by a cross-attention mechanism that dynamically integrates external conditions to guide the generation process. Furthermore, EMFusion integrates an imputation-based sampling strategy that treats forecasting as a structural inpainting task, ensuring temporal coherence even with irregular measurements. Unlike standard point forecasters, EMFusion generates empirical probabilistic prediction intervals from the learned conditional distribution, providing uncertainty-aware probabilistic forecasting rather than simple point estimation. Numerical experiments conducted on the multivariate narrow-band EMF datasets demonstrate that EMFusion with the contextual information of working hours outperforms the baseline models with or without conditions. The proposed EMFusion outperforms the best baseline by 23.85% in continuous ranked probability score (CRPS) and 13.93% in normalized root mean square error.