Designing an AI-driven digital twin architecture for building energy prediction

In the pursuit of global carbon neutrality, enhancing energy efficiency in buildings—particularly during their operational phases—has become a critical objective. This paper explores the potential of AI-driven Digital Twins in building operations and adopts a Design Science Research (DSR) methodology to guide the development, implementation, and evaluation of a data-driven solution. A six-layer conceptual architecture is proposed to support the integration of real-time data, AI models, and control logic. As a practical instantiation of this architecture, we present an application case involving three buildings, where energy consumption is predicted using machine learning. Several regression models are tested, with CatBoost Regressor achieving R 2 values above 0.92. The model is deployed as a Digital Shadow using the open-source platform Node-RED. In addition to energy forecasting, two simulated application cases demonstrate the architecture’s ability to support intelligent control strategies such as load shifting and anticipatory HVAC activation. The results validate the artifact’s effectiveness and highlight the potential of combining Artificial Intelligence and Digital Twins to improve building performance and sustainability through flexible, cost-effective, and open-source tools. • Six-layer architecture for energy forecasting and control in smart buildings. • Open-source proof-of-concept using Node-RED and edge computing. • Three application cases demonstrate forecasting and simulated control. • CatBoost model achieves R 2 > 0.92 across three diverse buildings. • Accessible solution for SMEs using low-cost hardware and free software.