Federated AI Learns Building Physics to Predict Indoor Temperatures Without Sharing Data

Buildings consume roughly 28 percent of the world’s energy, and a large share of that goes into keeping indoor spaces comfortable through heating, ventilation, and air-conditioning. A research team at Inha University in Incheon, Korea, has now reported a way to make the control of those systems smarter by teaching a single artificial intelligence model to predict indoor temperatures across several buildings at once, without any building ever handing over its raw sensor data. The study, published on 28 October 2026 in the journal Architectural Research, was conducted by Abu Talib, Kwangwon Choi, and Jaewan joe and describes a physics-informed neural network trained through federated learning, a technique in which multiple participants collaboratively build a shared model while keeping their own data locked away locally.

The motivation behind the work lies in a long-standing bottleneck for advanced building control. Model predictive control, which optimizes HVAC operation by forecasting how a building will respond to different control actions, is widely seen as an effective route to energy savings, but its performance depends entirely on the availability of an accurate and generalizable model of the building’s thermal dynamics. Developing such a model for each building individually can be time-consuming and expensive, sometimes accounting for a substantial portion of a project’s overall cost. Existing machine learning approaches usually treat each building in isolation, and while transfer learning has been explored as a way of moving knowledge from data-rich buildings to data-poor ones, its practical use is constrained by data availability, domain mismatches, and privacy concerns, since building owners are often reluctant to share operational data.

The Korean team’s answer was to combine two ideas that had rarely been brought together so directly. The first is federated learning, in which several buildings act as clients that train a model locally on their own measurements and send only the resulting parameter updates to a central server, which aggregates them into a global model. The second is physics-informed neural network modeling, in which physical knowledge is embedded into the training process through a custom loss function rather than relying purely on data fitting. In this study, the physical component was a resistance-capacitance thermal network, a simplified grey-box representation of building heat flows in which thermal resistances describe heat exchange between the outdoor environment, the building envelope, the room air, and adjacent zones, while thermal capacitances describe the energy stored in the envelope and the indoor air.

Crucially, rather than feeding each local RC model directly into training, the researchers first constructed a global RC model. They developed a local RC model for each of the four target zones, optimized its resistance and capacitance parameters against measured data using nonlinear optimization, then normalized the parameters, with resistances normalized by area and capacitances by thermal volume, and aggregated them across the participating zones. This generalized global RC model was then incorporated into the neural-network loss function as a physics-based constraint. The loss function combined two mean-squared-error terms, one measuring the difference between the network’s prediction and the measured room temperature, and the other measuring the difference between the network’s prediction and the temperature predicted by the global RC model, so that the network was trained simultaneously to fit the data and to remain consistent with the underlying thermal physics.

The federated training itself followed a FedProx-style procedure with a proximal regularization term that penalizes deviation of local model parameters from the global model, a safeguard against client drift when different buildings have different statistical distributions of temperature, solar gains, and HVAC operation. Each client trained a model with the same architecture, a sequence input layer followed by two stacked long short-term memory layers with layer normalization and dropout, a fully connected layer of 32 neurons with ReLU activation, and a single-neuron output for the next-step temperature. Inputs covered a three-hour historical window at 15-minute intervals, twelve time steps in all, including indoor and outdoor temperatures, solar heat gain, lighting heat gain, internal heat gain, HVAC cooling input, and corridor temperature. After local training, the server aggregated updates using sample-size-weighted federated averaging, so clients with more training sequences contributed proportionally more to the global model, and the process ran for 200 communication rounds.

The experimental data came from four zones in two campus buildings at Inha University: three zones on the fourth floor of a six-story building completed in 1954, each roughly 27 square meters with a ceiling height of 2.75 meters, and one zone in a second building of about 24 square meters. All zones were conditioned by direct-expansion air-conditioning systems, and the HVAC units were deliberately cycled on and off at intervals of roughly three to four hours to reduce the correlation between disturbances and control inputs that occurs during typical operation. The full dataset comprised 1,536 time-step observations at 15-minute intervals with seven input features per observation, split chronologically into two days of estimation data per zone and two days reserved for validation scenarios, a partitioning designed to prevent information leakage between training and evaluation.

The results showed that the global model achieved good single-step-ahead prediction performance across all four zones, with root-mean-square errors ranging from 0.36 degrees Celsius to 0.59 degrees Celsius and mean absolute errors from 0.25 to 0.57 degrees Celsius. The coefficient of determination reached 0.965, 0.979, 0.918, and 0.897 for the four zones respectively. For multi-step-ahead prediction, performed recursively over a 24-hour horizon of 96 consecutive 15-minute steps, the average RMSE was approximately 0.68 degrees Celsius during the estimation period, rising to approximately 0.99 degrees Celsius during validation. The increase reflects error accumulation inherent to recursive prediction, in which each predicted temperature is fed back into the input sequence for the next step, allowing early errors to propagate through the horizon, as well as the challenge of predicting conditions that differ from those seen during estimation.

Perhaps the most telling comparison came when the model was tested on a building it had never seen during training. In a leave-one-client-out validation, the zone from the second building was excluded from federated training and then evaluated as an unseen client using two days of data. The proposed physics-informed federated model achieved RMSE values of 0.74 degrees Celsius and 0.86 degrees Celsius for one-step-ahead and multi-step-ahead predictions respectively, compared with 0.86 and 0.98 degrees Celsius for a conventional purely data-driven federated model. During the validation period more broadly, the physics-informed model achieved RMSE values of 1.03 and 1.16 degrees Celsius for one-step and multi-step prediction, slightly better than the 1.05 and 1.20 degrees Celsius of the data-driven counterpart. Interestingly, the conventional data-driven model was actually more accurate during the estimation period, with average RMSE values of 0.32 and 0.38 degrees Celsius against 0.42 and 0.53 degrees Celsius for the physics-informed version, suggesting that the physics constraint acts as a regularizer whose benefit emerges most clearly when the model faces data it has not encountered before.

The study also quantified how much the physics-based inputs matter. When internal and lighting heat-gain inputs were excluded from the model, the multi-step-ahead validation RMSE rose from 0.99 degrees Celsius to 1.34 degrees Celsius, an increase of approximately 35.4 percent, while the estimation RMSE increased from 0.59 to 1.00 degrees Celsius, roughly 69.5 percent. For the global RC model alone, excluding the heat-gain inputs increased the average RMSE by about 6.5 percent. These figures underline that internal and lighting heat gains have a measurable influence on predictive accuracy and that omitting them can degrade both model performance and generalization.

The authors are careful to frame these findings as preliminary. The estimation and validation periods were short, which may restrict the model’s ability to capture seasonal variations and diverse operating conditions, and the participating clients were limited to four zones in two university buildings, three of them in the same building. The thermal and operational characteristics represented in the dataset may not fully capture the variability across different building types and HVAC systems. Moreover, in multi-step-ahead prediction the future disturbances and HVAC control variables were taken from measured data, whereas in practical model predictive control applications future weather must be forecast and control inputs must be determined by an optimizer over the prediction horizon. The team plans to investigate the effect of dataset size and temporal coverage, to validate the framework on larger and more diverse buildings, and to explore control strategies including set-point trajectories with pre-cooling scenarios and reinforcement learning-based HVAC control using the global model as a simulation environment. Even so, the demonstration that a shared, physics-consistent model can be built across buildings without pooling private operational data marks a meaningful step toward scalable, privacy-preserving smart building control.

Subject of Research: Physics-informed federated learning for indoor temperature prediction across building zones

Article Title: Physics-Informed Global Model Development via Federated Learning for Indoor Temperature Prediction Toward Advanced Building Control

Article References: Talib, A., Choi, K., & joe, J. (2026). Physics-Informed Global Model Development via Federated Learning for Indoor Temperature Prediction Toward Advanced Building Control. Architectural Research, 28(1), Article 26. https://doi.org/10.1007/s44515-026-00031-z

Image Credits: AI Generated

DOI: 10.1007/s44515-026-00031-z

Keywords: federated learning, physics-informed neural network, indoor temperature prediction, building energy efficiency, model predictive control, HVAC control, resistance-capacitance model, grey-box modeling, data privacy, smart buildings, LSTM networks, Inha University