Modeling and Assessment of RSSI in LoRa WAN Network Using a Neural Network Approach

This study employed a l0-fold cross-validation approach to train and validate neural networks to predict the RSSI (received signal strength indicator) in LoRaWAN communication. The model's performance was accurately estimated by averaging results across multiple validation sets and training sequences, and biased outcomes were mitigated. Implementing data normalization techniques, a simple neural network architecture, regularization approaches, and iteration limits ensured robustness within the limitations of a small dataset. The assessment of three neural network models using different optimization algorithms demonstrated their potential for accurate RSSI predictions, particularly for low-power IoT devices like LoRaWAN end devices, where energy efficiency is crucial. Lightweight neural networks on such devices enable efficient data processing while conserving energy, extending their lifespan, and reducing maintenance needs. Experimentation with deeper structures and unconventional activation functions may be necessary for complex problems, further enhancing performance in specific domains like RSSI assessment for LoRaWAN end devices.