Data-driven single gateway location estimation using improved channel model and LSTM

The Internet of Things (IoT) has expanded rapidly, with outdoor applications relying on Low-Power Wide-Area Networks (LPWANs) to deliver long-range, low-power connectivity. Among these, LoRaWAN enables energy-constrained sensors to communicate across large areas using minimal infrastructure. However, outdoor localization with LoRaWAN remains challenging: RSSI-based methods suffer from fading and noise, while advanced techniques such as TDOA and AoA require multiple synchronized gateways or specialized antennas, increasing cost and complexity. In this work, we propose a hybrid framework that integrates an improved empirical path-loss model with machine learning for single-gateway localization. First, we refine the Birmingham path-loss model using clustering and distribution-aware regression, improving prediction accuracy by up to 28.6% compared to regression-only baselines. Second, we develop a data-driven LSTM network that leverages sequential RSSI and SNR traces, guided by the improved channel model, to predict the sector location of mobile nodes. We further benchmark our proposed LSTM model against a CNN–LSTM baseline, a well-established deep learning paradigm for time-series classification. Experimental results highlight the superiority of our approach, achieving 72% accuracy compared to 65% for the CNN–LSTM baseline. By avoiding multi-gateway synchronization and heavy infrastructure, our approach demonstrates a practical, low-cost solution for single-gateway localization in dense urban environments. Experimental results in Birmingham city center highlight the robustness and feasibility of this method, making it suitable for smart city deployments, industrial monitoring, and other resource-constrained IoT applications.