PATH LOSS PREDICTION FOR MESH NETWORKS IN A REAL URBAN ENVIRONMENT USING MACHINE LEARNING

We investigate communication data extracted from two mesh networks in the metropolitan region of São Paulo, Brazil, owned by a company specializing in automating electric power distribution networks. We propose a methodology, based on machine learning frameworks, to help the planning and deployment of future mesh networks. Empirical and classical models were both used to predict the received signal strength indicator (RSSI). Besides, we employed Random Forest and Support Vector Regression considering the distance between the transmitter and receiver, information about obstruction in the first Fresnel zone, and terrain variability measures as features. According to the results, the Random Forest obtained the best performance among all methods. We also show the importance of each feature in the RSSI prediction and contrast the predicted results obtained from machine learning algorithms with the empirical and classical models.