The building sector, responsible for nearly 40% of global energy consumption, faces increasing pressure to improve efficiency particularly during the operational phase, which accounts for the majority of a building’s lifecycle energy use. Simulation...
The building sector, responsible for nearly 40% of global energy consumption, faces increasing pressure to improve efficiency particularly during the operational phase, which accounts for the majority of a building’s lifecycle energy use. Simulation-based decision-making is essential for improving operational performance of building systems. By quantitatively evaluating control strategies, simulation models help identify energy-saving options and inform real-time decisions. However, despite its potential, simulation-based decision-making is rarely applied in practice, mainly due to the difficulty of developing models that are both reliable and applicable under real-world conditions.
For simulation models to be truly useful in building operations, they must satisfy two critical requirements: causality and practicality. A practical model should be trainable with limited and often fragmented real-world data, while a model with strong extrapolation ability must provide reliable predictions even under unseen scenarios. Without these traits, simulation models risk offering misleading guidance, especially in dynamic and data-scarce environments common to building operations.
To address these limitations, this study investigates how simulation models can be enhanced through bi-directional knowledge share. Rather than developing models in isolation, I propose enabling communication between models trained on different but related datasets. This approach allows models to mutually expand their understanding of system behavior, even in the presence of data imbalance or domain gaps. To implement this collaborative learning framework, federated learning (FL) is adopted, which facilitates parameter-level knowledge exchange while preserving data privacy and ownership across different domains.
Beyond improving model development, the study also addresses the question of how to evaluate whether a model is “good enough” for decision-making. We focus on the causal knowledge embedded in the model, specifically, the causal relationships between control variables and system outcomes. Using double machine learning (DML), we extract and quantify these causal effects and introduce uncertainty in causal inference as a new evaluation metric. Unlike conventional accuracy-based metrics that focus only on in-distribution prediction error, uncertainty in causal inference captures the stability and reliability of a model’s predictions under unseen conditions, offering a more in-depth measure of extrapolation ability.
The proposed approach is validated through simulation-based case studies involving various building types and HVAC systems, including chillers, variable refrigerant flow (VRF) systems, and packaged terminal air conditioning (PTAC) units. Results show that federated learning significantly enhances the reliability and extrapolation performance of simulation models, particularly when knowledge is shared with similar dynamic characteristics. In addition, the results demonstrate that uncertainty in causal inference can serve as a meaningful indicator of a model’s extrapolation ability. As uncertainty in causal inference decreases, the model's predictions become more stable and physically plausible in unobserved scenarios, suggesting that this metric can be used to assess and infer improvements in performance beyond conventional accuracy-based evaluations.
In summary, this study proposes a novel methodology to enhance the causality and extrapolation ability of simulation models for building HVAC operations. It addresses the challenges of limited data and model robustness by adopting a federated learning framework, enabling decentralized knowledge sharing across systems. At the same time, it introduces causal inference techniques as a way to evaluate model reliability under unobserved conditions. Through this integrated approach, the study offers two key perspective expansions: (1) a shift from isolated to federated frameworks for simulation model development, and (2) a shift from traditional error-based evaluation to causality-based evaluation focused on the stability of inferred causal relationships. These contributions lay the groundwork for more reliable and informed decision-making in real-world building operations.