Spatio-Temporal Model for Predicting Multivariate Weather-Induced Attenuation in Wireless Networks

This study introduces an innovative approach for predicting weather-induced attenuation in wireless communication networks (WCNs) that are highly sensitive to environmental changes. We aim to accurately predict short-term signal levels for both communication and sensing applications, a capability that will be crucial as next-generation networks (NGNs) operating in high-frequency millimeter-wave (mmWave) bands unlock advanced sensing opportunities. Our framework leverages a multivariate model to account for the influence of dynamic weather conditions on multiple links in communication networks. We present a selective bidirectional spatio-temporal network (S-BSTN), augmented with dual-attention mechanisms to effectively capture spatial and temporal dynamics across multiple signal levels for predictive tasks. Real-world experiments on operational networks demonstrate that our methodology achieves over 20% improvement in RMSE across diverse network conditions, consistently outperforming state-of-the-art prediction models. Our research utilizes wireless communication data to monitor, track, and predict weather-induced phenomena, transforming these data into a tool for predictive modeling. By leveraging sub-minute temporal resolution and high spatial density, we demonstrate the potential to generate rainfall rate maps and transform communication networks into highly effective, high-resolution sensors.