Louis Closson

Abstract:
Optimizing energy consumption and enabling smart building management are key challenges in response to today’s environmental, economic, and social concerns. Buildings, which account for 34% of global energy consumption and 37% of CO₂ emissions, are central to sustainability efforts. This thesis explores the potential of machine learning and DataOps practices to support effective smart building management, leveraging Internet of Things (IoT) infrastructures and advanced data analysis methodologies.

This work contributes in several ways. First, a detailed dataset was developed using the GreEn-Er building’s API, enabling the analysis of real time series data, such as temperatures and CO₂ levels, for anomaly detection and clustering. A second dataset was created by equipping a commercial building with sensors, providing open access to long-term data on temperature, air quality, and HVAC system consumption. The thesis also investigates predictive modeling techniques to forecast air quality, temperature, and energy consumption. A key innovation lies in the adaptation of transfer learning approaches to enhance algorithm performance on diverse sensor data while minimizing computational costs. The methods and results, validated through real-world implementations, offer valuable insights for integrating artificial intelligence into smart building management. By improving anomaly detection, prediction accuracy, and transfer learning capabilities, this research addresses the growing need for sustainable and scalable building solutions, paving the way for future advances in industrial applications. 

Keywords:
Deep learning, time series, data forecasting, low-power sensor networks (LPWAN), smart building