SmartBuildSim: An Open-Source Synthetic-Twin Framework for Reproducible AI Benchmarking in Smart-Building Analytics

Area of Science:

• Building energy systems
• Artificial intelligence in smart buildings
• Data science and simulation

Background:

• High-fidelity building simulators are computationally expensive.
• Existing synthetic datasets often lack realistic temporal dynamics and reproducibility.
• AI research in smart buildings requires reliable and scalable testbeds.

Purpose of the Study:

• Introduce SmartBuildSim, an open-source synthetic twin framework.
• Enable generation of configurable and reproducible multi-sensor building data streams.
• Provide a lightweight alternative to complex physical digital twins for AI research.

Main Methods:

• Utilized lightweight statistical models with tunable parameters (trend, seasonality, correlation, delays, anomalies).
• Implemented deterministic seeding for experiment-level reproducibility.
• Developed modular pipelines for unified evaluation across forecasting, anomaly detection, and reinforcement learning (RL).

Main Results:

• Validated synthetic data against ASHRAE reference signals, showing realistic magnitude and variability (KS ≈ 0.32; DTW ≈ 9.69).
• Demonstrated strong forecasting performance with linear models (RMSE ≈ 21.27).
• Showcased superior anomaly detection (IsolationForest vs. LOF: F1 ≈ 0.17 vs. 0.10) and RL convergence (Soft-Q Learning variance reduced by >95%).

Conclusions:

• SmartBuildSim offers a transparent, lightweight, and reproducible alternative to high-fidelity simulators.
• The framework effectively bridges the gap between simple synthetic data and complex physical digital twins.
• Provides a practical testbed for advancing AI research in smart buildings.