Abstract
The rising demand for fast, reliable communication networks has resulted in greater dependence on microwave and millimeter-wave radio frequency links spectrum. However, radio communication systems utilizing spectrum above 10 GHz face significant performance challenges due to attenuation caused by atmospheric effects; this can often lead to service disruptions and severe link outages. Rainfall is a major meteorological factor that impacts signal attenuation in both terrestrial communication systems and satellite-to-Earth station links. Traditional empirical models, involving the International Telecommunication Union (ITU-R) and certain probabilistic modelling techniques, have been widely used to estimate rainfall-induced attenuation. However, these models rely on simplified statistical assumptions and often fall short in accurately capturing the intricate and nonlinear relationships among rainfall microparameters, particularly the raindrop size distribution (DSD) and intensity of the rainfall, which are embedded within extensive rainfall data.
To address these limitations, this research explores the use of supervised learning techniques for modelling DSD and predicting rainfall attenuation over radio links. Three machine learning regression models, the random forest (RF), 𝑘-nearest neighbours (KNN) and decision trees (DT) are employed to develop three DSD models, trained on disdrometer data collected in Durban, South Africa (29.8651°S, 30.9734°E) between January 2018 and December 2019. The dataset, collected at one-minute intervals, is categorized into four distinct rainfall regimes according to rainfall intensity: drizzle, widespread rainfall, showers, and convective thunderstorms.
The results indicate that machine learning models surpass traditional statistical methods in predicting DSD and estimating rainfall attenuation. Among them, KNN achieves the highest accuracy, particularly under thunderstorm conditions, with an average coefficient of determination (𝑅² Score) of 0.92 across all rainfall regimes. While RF also performs well, the Decision Tree model shows relatively lower predictive accuracy, with an average root mean square error (RMSE) of 8.92 and a mean absolute error (MAE) that is 4.15 units higher than those of the RF model. Additionally, an adaptive optimal 𝑘-selection method for KNN is introduced, dynamically tuning hyperparameters to improve prediction accuracy across varying rainfall conditions. This novel approach enhances model adaptability and precision in estimating rainfall-induced attenuation.
vi
Comparative analysis reveals that the predicted rainfall attenuation at a 0.01% rain rate exceedance closely matches observed measurements across annual, seasonal, and monthly periods. The proposed models are further validated through comparisons with ITU-R models and actual link measurements, demonstrating a strong correlation with real-world data.
By showcasing the enhanced accuracy and adaptability of ML-based models, this study provides valuable insights, confirming their effectiveness for link budgeting and network planning in high-rainfall environments. By leveraging data-driven models, this research provides a more robust approach to rainfall attenuation estimation, improving network resilience, optimizing link budgeting, and enhancing communication system reliability in regions prone to heavy rainfall. These findings are particularly relevant for the deployment of high-frequency communication systems, including 5G and future 6G networks.