⚡ electrical engineering

AI-Assisted Link Quality Prediction for Weather-Aware Outdoor Terahertz Communications: Forecasting, Uncertainty, and Explainability Analysis

This paper presents a real-weather-driven framework for forecasting outdoor terahertz link quality in Bangladesh using six years of meteorological data, demonstrating that while AI models outperform persistence baselines, they exhibit significant errors and overconfidence specifically during rain-induced degradation, a finding validated across two distinct climatic sites.

Original authors: Israt Jahan, Md. Jahedul Islam

Published 2026-09-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Israt Jahan, Md. Jahedul Islam

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 data moves at the speed of light, carrying terabytes of information in a single blink, enabling the ultra-fast connections promised by the next generation of wireless networks. This is the promise of terahertz communication, a technology that operates on frequencies far higher than those used by current smartphones or Wi-Fi. While these high frequencies offer a vast highway for data, they come with a significant vulnerability: the air itself. Unlike lower frequencies that can punch through clouds or light rain, terahertz signals are extremely sensitive to the atmosphere. Even a light drizzle can scatter the signal, and a heavy downpour can sever the connection entirely, turning a high-speed link into a dead line within meters. For engineers building these networks, the challenge is not just building the hardware, but predicting when the weather will break the link before it happens, allowing the system to adapt instantly.

Researchers at the Chittagong University of Engineering and Technology in Bangladesh have tackled this problem by building a digital twin of a real-world terahertz link and teaching a computer to forecast its fate. Instead of guessing how a link might behave under a fixed, theoretical rainstorm, they fed a physics-based simulator six years of actual, hour-by-hour weather data from the city of Sylhet. This simulator calculated exactly how much signal would be lost to rain, humidity, and the natural drift of the equipment, translating those losses into a single, easy-to-read score called the Link Quality Index. This score, which ranges from zero to one, tells a network operator how well the connection is performing at any given moment, combining signal strength, error rates, and hardware stability into one number. The goal was to see if artificial intelligence could look at the past weather and the current link quality to predict what the score would be one, three, six, or even twelve hours into the future.

The team tested seven different types of prediction models, ranging from simple statistical methods to complex deep-learning networks that mimic the human brain. They found that the most effective approach was surprisingly straightforward. A model based on gradient-boosted trees, a type of algorithm that builds a decision-making structure from many simple rules, outperformed the more complex deep-learning systems. In fact, the complex models did not offer a significant advantage in accuracy. The best models could predict the future link quality with a skill that was about 21% to 31% better than simply guessing that the weather would stay the same as it is right now. This improvement grew larger the further into the future the forecast looked, because the "guess it stays the same" method fails quickly as weather changes, while the trained models learned the patterns of how rain and humidity evolve over time.

However, the study revealed a critical limitation that aggregate numbers often hide. While the models performed well on average, they struggled significantly when the link was already in bad shape. During the most severe rain events, when the link quality dropped to its lowest levels, the models' predictions became unreliable, often performing worse than a simple average. This is the exact scenario where a prediction is most needed, yet the system falters. To address this, the researchers reframed the problem: instead of trying to predict the exact quality score, they asked the model to simply warn if the link was about to drop below a critical threshold. In this binary mode of "warning or no warning," the models remained effective, successfully identifying degraded conditions much better than a basic guess, even when the exact numbers were off.

The researchers also explored how much of the prediction came from the weather itself versus the history of the link. They discovered that the most powerful predictor was simply the link's own recent performance; knowing how the link behaved an hour ago was more important than knowing the current rain rate. Surprisingly, adding detailed historical weather data or complex time-series analysis provided only a tiny boost in accuracy. This suggests that for this specific type of short-range link, the current state of the weather and the immediate past of the connection are the dominant factors. To ensure these findings were not just a fluke of one location, the team repeated the entire experiment in Rajshahi, a city with a much drier climate. The results held true: the simple model remained the best performer, and the struggle to predict the worst-case scenarios persisted, confirming that these are fundamental challenges of the technology rather than quirks of a single dataset.

Ultimately, this work provides a clear roadmap for how to manage future wireless networks in a changing climate. It demonstrates that while we cannot yet perfectly predict the exact moment a terahertz link will fail during a storm, we can build systems that reliably warn us when trouble is coming. The study emphasizes that for this specific engineering problem, simpler, faster, and more interpretable models are often superior to complex, black-box deep learning systems. By combining physics-based simulations with machine learning, the researchers have created a tool that can help network operators anticipate weather-driven outages, ensuring that the next generation of ultra-fast communication remains robust even when the sky turns gray.

Technical Summary: AI-Assisted Link Quality Prediction for Weather-Aware Outdoor Terahertz Communications

Problem Statement
Outdoor Terahertz (THz) communication links, a candidate physical-layer technology for 6G, suffer severe degradation from rain and atmospheric attenuation. While existing literature characterizes these impairments using static, fixed rain-rate scenarios, there is a lack of predictive frameworks that forecast weather-driven link degradation as a dynamic process using real meteorological time series. Furthermore, prior work often focuses on isolated metrics like Bit Error Rate (BER) or Signal-to-Noise Ratio (SNR) without coupling them into a composite, application-relevant metric, nor do they typically integrate rigorous uncertainty quantification and explainability within a single, leak-free forecasting pipeline.

Methodology
The authors propose a comprehensive forecasting framework driven by six years of hourly meteorological data (51,888 records) from Sylhet, Bangladesh (2020–2025). The methodology consists of three primary stages:

Physics-Based Simulation and Link Quality Index (LQI) Definition:
- A simulator drives channel conditions using real-time rain rate, humidity, temperature, and pressure. It employs ITU-R P.838-3 for rain attenuation, ITU-R P.676 for gaseous attenuation, and a modified Friis link budget for a 300 GHz, 200-meter link.
- A composite Link Quality Index (LQI) is defined as a weighted sum of three normalized margins: Channel Margin (received power), Reliability Margin (QPSK BER), and Hardware Margin (pointing loss). The weights are fixed at 0.45, 0.40, and 0.15, respectively. This metric is distinct from raw BER/SNR as it reflects the link's ability to support applications.
- The dataset includes adaptive modulation logic, where the system selects the highest-order modulation (up to 16-QAM) satisfying a target BER of $10^{-6}$ , though QPSK dominates the observed hours.
Forecasting Pipeline and Feature Engineering:
- Data Split: A strict, leak-free chronological split (70% training, 15% validation, 15% testing) is used to prevent data leakage.
- Features: Inputs include instantaneous weather, cyclical time encodings, weather history (lags and rolling statistics), and LQI history.
- Models: Seven models are benchmarked across four forecast horizons (1, 3, 6, and 12 hours): Persistence, Linear Regression, Random Forest, XGBoost, GRU, LSTM, and a lightweight Transformer.
- Evaluation: Metrics include RMSE, MAE, $R^2$ , and forecast skill relative to persistence. Robustness is assessed via multi-seed testing and the Diebold–Mariano significance test.
Uncertainty and Explainability:
- Uncertainty: Monte Carlo (MC) Dropout is applied to the GRU model to generate predictive intervals and assess calibration.
- Explainability: SHAP (SHapley Additive exPlanations) is used to attribute predictions to specific features.
- Ablation: Feature-group ablation studies isolate the value of instantaneous weather versus historical data.
- Stratified Evaluation: Performance is analyzed specifically under degraded conditions (rain events and low LQI quartiles) to reveal failures masked by aggregate metrics.

Key Contributions

Real-Weather-Driven Dataset: A reusable, six-year, hourly-resolution dataset linking real meteorological observations to a physics-based THz channel model, moving beyond fixed-scenario characterizations.
Composite LQI Metric: A physically anchored, normalized metric combining channel, reliability, and hardware margins, with a documented sensitivity analysis confirming its robustness to weight variations.
Rigorous Benchmarking: A seven-model comparison under a strictly leak-free protocol, including a Transformer baseline, supported by hyperparameter tuning, multi-seed robustness, and statistical significance testing.
Uncertainty and Explainability Integration: An honest assessment of model calibration (noting overconfidence at longer horizons) and feature attribution via SHAP.
Cross-Site Validation: Application of the identical pipeline to a second, climatically contrasting site (Rajshahi, Bangladesh) to verify generalizability.

Results

Forecast Skill: All learned models outperform the persistence baseline, with skill increasing from ~21.5% at 1-hour to ~31.4% at 12-hour horizons. XGBoost and GRU are the top performers, often statistically indistinguishable at short-to-medium horizons, though XGBoost shows a statistically significant advantage at the 12-hour horizon in the Sylhet dataset.
Model Complexity vs. Performance: The Transformer baseline does not outperform simpler models (XGBoost, GRU, LSTM) at any horizon. Feature ablation reveals that instantaneous weather data alone (8 features) captures the majority of achievable skill; adding historical data yields only marginal gains (0–1.5 percentage points).
Uncertainty Calibration: MC Dropout intervals are reasonably well-calibrated at short horizons (~~93% coverage) but become progressively overconfident as the horizon extends (~~90% coverage at 12 hours), despite rising RMSE.
Degraded Condition Performance: Aggregate metrics mask severe failures. In the worst-LQI quartile (rain-affected conditions), the regression model performs worse than a simple mean predictor ( $R^2$ becomes negative). However, when reframed as a binary degradation detection task (predicting if LQI drops below a threshold), the model significantly outperforms persistence across all horizons.
Cross-Site Consistency: The central findings, including the dominance of instantaneous weather and the failure of aggregate metrics to capture worst-case errors, hold true for the drier Rajshahi site, suggesting these are properties of the LQI-forecasting problem rather than site-specific artifacts.

Significance and Claims
The paper claims its primary contribution is methodological rigor and system-level integration rather than the invention of a new learning algorithm. It demonstrates that for this specific problem scale (outdoor THz links in a specific climate), established, lightweight models (like XGBoost) are sufficient for point forecasting and are preferable for deployment due to lower computational cost.

The authors emphasize that the inclusion of deep learning models (GRU, LSTM, Transformer) serves two purposes: establishing a fair comparison to prove that complex temporal modeling offers no marginal gain in point forecasting for this dataset, and enabling uncertainty quantification (via MC Dropout), which gradient-boosted trees do not natively support.

Crucially, the paper highlights a critical limitation: while aggregate metrics suggest good performance, they substantially understate forecast errors during the very rain/degraded conditions that motivate the research. The study concludes that while exact-value forecasting fails in worst-case scenarios, the framework successfully provides a reliable binary early-warning signal for downstream controllers. The authors position this work as a foundation for future field validation and adaptive resource allocation, noting that the current study is simulation-based and relies on standardized propagation models rather than direct field measurements.

Technical Summary: AI-Assisted Link Quality Prediction for Weather-Aware Outdoor Terahertz Communications

More like this