Mobility-Aware Temporal Machine Learning for Proactive Gateway Selection in Industrial LoRaWAN Networks

Abstract

 Reliable gateway selection in mobile LoRaWAN deployments is challenged by rapid RSSI fluctuations caused by mobility, multipath propagation, and dense urban interference. This paper introduces a temporal machine-learning framework that performs \emph{proactive gateway forecasting} using short RSSI/SNR sequences, enabling prediction of the most reliable gateway for the next uplink rather than relying on instantaneous measurements. Experiments on two heterogeneous real-world datasets—an industrial mobile trace and a city-wide deployment with forty gateways—show that gradient-boosted trees, BiLSTM models, and lightweight hybrid ensembles significantly improve fairness (Macro-F1) and deliver consistently high Top-$k$ recall, allowing the network server to shortlist only a small set of candidate gateways with minimal loss in reliability. The method integrates into a two-stage architecture in which the ML model functions as a Stage-1 gateway-ranking module, after which existing LoRaWAN network servers may apply duty-cycle, queue-aware, and policy constraints without modification. SHAP analysis verifies that the learned predictors align with wireless-channel behaviour, emphasising features such as mean RSSI, short-term variance, and dominance gap. Overall, the results demonstrate that temporal forecasting provides a practical, interpretable, and scalable foundation for mobility-aware gateway selection in real LoRaWAN networks.