RSSI prediction for improved LoRa communications performance

Abstract

Optimal tuning of LoRa communications is of great importance to optimize the efficiency and robustness of LoRa networks. This optimization is necessary to enhance the performance of IoT communications in industrial environments, enabling the full benefits of LoRa technology to be realized in sometimes difficult contexts. This paper focuses on the prediction of RSSI as a key indicator of LoRa performance based on weather data, SNR, and the distance between nodes and the gateway in an outdoor environment. The aim is to evaluate seven machine learning models (Decision Trees, Linear Regression, Process Gaussian, Random Forest, SVM, Gradient Boosting, and AdaBoost) for RSSI prediction, with the aim of subsequently optimizing LoRa performance by adjusting its specific parameters, such as spreading factor, transmission power and bandwidth, and coding rate. The results show that three models showed promising performance, each adapted to a specific context. The Gaussian Process model with the “ardexponential” kernel is particularly well adapted when accuracy is paramount, offering the best prediction with an RMSE of 0.0233 and an efficiency of 95.7261%. For situations requiring a balance between accuracy and temporal efficiency, the Random Forest model demonstrates robust performance with an RMSE of 0.0323, a training time of 114.7459s, and a prediction time of 5.8427s. Finally, in a context focused on temporal efficiency while maintaining an acceptable level of accuracy, the decision tree model is considered appropriate.