Decentralized Intelligence in SBs: Federated LSTM-Powered Digital Twins for Sustainability

Prabhu RajaramGnana Swathika OV^*

• Vellore Institute of Technology, Chennai, Chennai, India

The growing demand for intelligent, data-driven building automation poses significant challenges related to data privacy, scalability, and heterogeneity across distributed environments. Traditional centralized machine learning (ML) approaches often require aggregating sensitive sensor data to a central server, raising privacy concerns and limiting real-time responsiveness. To address these limitations, this study proposes a novel framework that integrates federated learning (FL) and digital twin (DT) technologies for privacy-preserving, real-time occupancy detection in smart building (SB). Leveraging time-series data from environmental sensors, a long short-term memory (LSTM) model is collaboratively trained across distributed clients using the federated averaging (FedAvg) algorithm, ensuring data remains local to the source. A personalized fine-tuning phase enhances local model adaptation, making the system robust to client-specific variations and non-IID data. The resulting model is deployed within a streamlit-based digital twin interface that enables real-time visualization of occupancy states, sensor dynamics and model predictions, including rolling forecasts, confidence levels and error diagnostics. This integrated approach not only enhances privacy and adaptability but also enables proactive energy management and operational transparency. By unifying federated sequence learning with dynamic digital representations, the framework contributes to the development of scalable, secure and sustainability-aware SB systems.