Evaluating Machine Learning Classifiers for Prediction in an IoT-based Smart Building System

Abstract

The integration of Machine Learning (ML) with the Internet of Things (IoT) allows effectively analyzing the huge amount of gathered data, thus making them more meaningful and helping to accurately identify anomalies and potential problems. This paper presents the use of ML techniques in the analysis of data gathered from Smart Place, a real-world IoT-based smart building system that automatically controls air conditioners aiming at saving energy. A predictive agent uses these techniques to determine the actual state of air conditioners based on data about temperature, humidity, and the presence of people in the monitored environments. Four well-known ML classifiers (namely k-Nearest Neighbors, Multi-Layer Perception neural networks, Random Forest, and Support Vector Machines) were considered in an empirical study aimed to evaluate their suitability with respect to accuracy, resource utilization, and execution time. Obtained results showed a maximum average accuracy of 96.5% in the prediction of the state of air conditioners, besides the feasibility of using alternative models in compliance with resource constraints related to the IoT scenario.