Continuous learning digital twin for predictive building energy management

Digital Twins are increasingly being adopted to support the monitoring and optimization of buildings. Recent studies highlight both the growing importance of digital twins in urban and building contexts and the need for tighter integration with operational data workflows, more so with building energy data. However, many current implementations still rely on static or loosely connected analytical components, limiting their value for operational use and decision support. This restricts their usefulness for building energy management, where energy forecasting, subsystem-level interpretation, and prospective scenario evaluation are essential (Argota Sánchez-Vaquerizo, 2024; Arsecularatne et al., 2024; Boukaf et al., 2024).<br/>This research presents a sensor-driven digital twin framework for smart building energy management, demonstrated on the building of the Big Data for Smart Society (GATE) Institute in Sofia, Bulgaria. The proposed framework integrates four main components within a unified operational workflow: (1) circuit-level electricity measurements, which are electrical measurements collected at smart meters installed at individual circuits (2) data-driven energy forecasting using machine learning (ML)-based global and local forecasting models, (3) scenario simulation, where predicted electricity demand can be evaluated under different operational conditions and (4) a semantically enriched 3D building information model (BIM) that links energy measurements to building systems. Unlike many digital twin approaches that depend on static simulations, this approach allows continuously updated energy analysis driven by incoming sensor data.<br/>The digital twin framework operates using high-resolution sensor data extracted and maintained for the GATE building. The circuit-level electrical measurements are linked to a semantically enriched three-dimensional building information model (BIM) that integrates energy data with the corresponding building infrastructure. This data is used to train time-series prediction models integrated into the digital twin architecture, producing short- and medium-term forecasts to support scenario testing and iterative system refinement.<br/>A two-level modelling strategy is adopted. At the global level, a forecasting model captures the whole-building electricity consumption behaviour using circuit-level measurements, temporal variables, seasonality indicators, and external temperature as input features. At the local level, separate models are trained to predict electricity demand for individual circuits, enabling a more granular analysis. Candidate models are trained and evaluated by comparing predicted electricity demand against observed sensor measurements, with the best-performing model selected based on Mean Absolute Error and Root Mean Square Error (RMSE). The data is split into 80% training data, 5% validation data, and 15% testing dataset.<br/>Preliminary results demonstrate that for short-term forecasting, LSTM reduces forecasting error by 18%, compared to the baseline(persistence), and performs 57% better than ARIMA, a statistical model. To support continuous learning, the models are retrained on new sensor data, enabling the system to adapt to evolving usage patterns. This approach builds upon developments in smart building performance monitoring and predictive energy modelling (Al Dakheel et al., 2020; Misiurek et al., 2025). The main contribution of this work is a digital twin architecture that integrates circuit-level data, machine learning forecast models, and BIM representation to enable continuous predictive energy monitoring.