Abstract
System identification of smart buildings is necessary for their optimal control and application in demand response. The thermal response of a building around an operating point can be modeled using a network of interconnected resistors with capacitors at each node/zone called RC network. The development of the RC network involves two phases: obtaining the network topology, and estimating thermal resistances and capacitances. In this article, we present a provable method to reconstruct the interaction topology of thermal zones of a building solely from temperature measurements. We demonstrate that our learning algorithm accurately reconstructs the interaction topology for a 5-zone office building in EnergyPlus with real-world conditions. We show that our learning algorithm is able to recover the network structure in scenarios where prior research prove insufficient.
The Challenge of Smart Building Control
To optimize energy use in "smart buildings," engineers often use Model Predictive Control (MPC). However, MPC is only as good as the thermal model it relies on. The gold standard for these models is the RC (Resistor-Capacitor) network, where rooms are nodes (capacitors) and the walls/airflow between them are connections (resistors).
The challenge is that for many modern buildings—especially open-plan offices—the exact "topology" (the map of which zones affect which) is often unknown. While high-fidelity simulations like EnergyPlus exist, they are too computationally heavy for real-time control. Previous attempts to "learn" these maps from data often failed because they couldn't handle "colored noise"—real-world disturbances like fluctuating occupancy or weather that are temporally correlated.
Methodology: Topology Reconstruction via Wiener Filtering
The researchers propose a two-step approach based on Multivariate Wiener Filtering to map the "thermal wiring" of multi-zone buildings using nothing but temperature readings.
- Moral Graph Construction: They first use Wiener filters to identify a "moral graph." This graph includes all direct neighbors (rooms that share a boundary) but also "spurious" edges between two-hop neighbors (rooms separated by a third room).
- Phase-Based Pruning: This is the paper's core innovation. The authors mathematically prove that the phase response of a Wiener filter for a "strict two-hop neighbor" is exactly $\pi$ (180 degrees) across all frequencies. By checking the phase angle of the identified connections, the algorithm can "prune" the spurious links, leaving only the true physical connections.
To handle situations with limited data, they also introduced a regularized version of the algorithm that uses sparsity-based constraints to prevent overfitting.
Experimental Setup and Validation
The approach was validated using EnergyPlus simulations of a 5-zone office building model located in Minneapolis, MN. The simulation incorporated actual meteorological year data to simulate ambient conditions and synthetic noise to simulate internal heat loads from lights, occupants, and equipment.
Key Insights
- Magnitude is Deceiving: You cannot distinguish a direct neighbor from an indirect one just by looking at the "strength" of the thermal correlation. Some indirect connections appear stronger than direct ones in the data.
- Phase is the "Smoking Gun": The phase shift of $\pi$ is a robust physical signature of an indirect connection in an RC network. This allows for a "provable" reconstruction rather than a purely statistical guess.
- Handling Correlation: Unlike previous methods that required "white noise" (random, uncorrelated inputs), this method works with "colored noise," making it much more applicable to real buildings where heat loads follow patterns.
Results and Performance
In simulations, the algorithm achieved 0% error (perfect topology reconstruction) as the sample size increased. The proposed method significantly outperformed standard regression and Graphical Lasso (a common tool for structure learning). In some scenarios where the proposed method had near-zero error, standard regression had an error rate of over 50%. Furthermore, the regularized version of the algorithm showed much higher accuracy in "low sample regimes," making it practical for buildings where long-term historical data might not be available.
Conclusion and Future Work
While this research successfully addresses topology reconstruction, future work is needed to accurately estimate specific thermal resistance ($R$) and capacitance ($C$) values. Additionally, more complex dynamics, such as the thermal mass of internal walls, should be integrated into the model.
Beyond control optimization, the authors highlight a potential cyber-security risk: because building structures can be reconstructed solely from temperature data, an attacker could potentially "map" a secure facility just by hacking into the thermostat system.
Summary updated