🤖 AI

ADAPT: Physics-Aware Diffusion-based World Models for Adaptive Predictive Transferable HVAC Control

The paper proposes ADAPT, a physics-aware conditional diffusion world model that integrates learnable heat-balance regularizers to overcome thermodynamic inertia and partial observability challenges, enabling robust and transferable HVAC control that significantly reduces energy consumption and occupant discomfort across both standard and unseen climate conditions.

Original authors: Xu Yang, Kailai Sun, Dianyu Zhong, Qianchuan Zhao

Published 2026-08-21

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Xu Yang, Kailai Sun, Dianyu Zhong, Qianchuan Zhao

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Buildings are among the largest consumers of energy on the planet, responsible for roughly one-third of global electricity use and carbon emissions. A significant portion of this energy goes toward heating, ventilating, and air-conditioning systems, which work tirelessly to keep indoor spaces comfortable. However, managing these systems is surprisingly difficult. Unlike a light switch that responds instantly, a building has a kind of memory. Thick walls, concrete floors, and furniture absorb heat slowly and release it even more slowly. This means that when a thermostat is adjusted, the temperature inside does not change immediately; the effect of that action might not be felt for hours. Furthermore, the conditions outside are constantly shifting, with seasonal weather, sunlight, and the number of people in a room all changing the internal climate in complex ways. Because of these delays and changing conditions, traditional control systems often struggle to predict the future state of a room, leading to wasted energy or uncomfortable temperatures.

To solve this, researchers have developed a new approach called ADAPT, which acts as a predictive guide for building managers. Instead of simply reacting to the current temperature, this system learns to imagine what the building's temperature will be in the future if the current settings are kept the same. It uses a sophisticated type of artificial intelligence known as a diffusion model, which is excellent at generating realistic sequences of data, to create a "thermal baseline." Think of this baseline as a forecast of the building's natural behavior, showing how the heat will drift through the walls and air over time. By understanding this delayed response, the system can make smarter decisions, turning the heating or cooling on or off at the right moment to avoid overshooting the target temperature.

The researchers tested this idea in two different ways. First, they used a high-fidelity simulation of a real seven-zone office building in China, calibrated with thousands of hours of real-world data. Second, they tested it in a different environment using a global benchmark that simulated conditions in both the cold climate of Stockholm, Sweden, and the hot desert climate of Arizona, USA. In these tests, the new system was compared against standard control methods and other advanced artificial intelligence techniques. The results showed that by using this predictive guide, the system reduced energy consumption by 7.3 percent and significantly improved occupant comfort, cutting the time people spent in uncomfortable temperatures by 30.2 percent.

What makes this discovery particularly important is how well the system handles situations it has never seen before. Most computer models are trained on specific data, such as summer weather in one city, and they often fail when the season changes or when they are moved to a different climate. The researchers found that their system remained robust even when transferred from summer to winter or from a cold region to a hot one. This reliability comes from a unique feature of the design: the system is not just learning patterns from data; it is also guided by the fundamental laws of physics. It is taught to respect the basic rules of how heat moves between rooms, how sunlight warms a space, and how people generate body heat. This physical grounding prevents the system from making unrealistic guesses when the weather changes, ensuring that its predictions remain accurate even in unfamiliar conditions.

The study demonstrates that combining advanced artificial intelligence with the unchanging principles of physics can create a more efficient and adaptable way to manage our built environment. By giving the control system a clear picture of the future thermal state of a building, it can stop reacting blindly to the present moment and start planning for what is coming next. This approach offers a practical path toward reducing the energy footprint of buildings while keeping people comfortable, addressing a critical challenge in the global effort to mitigate climate change. The work suggests that the future of building management lies not in faster computers alone, but in smarter models that understand the physical world they are trying to control.

Technical Summary: ADAPT

Problem Statement

Building heating, ventilation, and air-conditioning (HVAC) systems are critical for urban energy efficiency and occupant comfort, yet they face significant control challenges. The core difficulties include:

Delayed and Partially Observable (PO) Dynamics: Due to thermal inertia, the effects of control actions (e.g., changing setpoints) are not immediately visible but emerge after multiple intervals. Furthermore, critical states like heat storage in structural mass are unmeasurable, rendering the system partially observable.
Distribution Shifts: Indoor thermal dynamics vary substantially across seasons, weather conditions, and climate regions. Existing data-driven methods often fail to generalize when operating conditions shift out-of-distribution (OOD), such as moving from summer to winter or between different climate zones.
Limitations of Current Methods: Model Predictive Control (MPC) is sensitive to model mismatch and requires manual calibration. Deep Reinforcement Learning (DRL) methods are often reactive, struggling with delayed credit assignment, and lack explicit modeling of latent thermal states, leading to poor generalization across different environments.

Methodology: ADAPT Framework

The authors propose ADAPT (A Diffusion-based Adaptive Physics-aware indoor environmental world model for Transferable HVAC control), a framework designed to learn a transferable Indoor Environmental World Model (IEWM) that guides downstream RL controllers.

1. Generative IEWM via Conditional Diffusion

The core of ADAPT is a conditional diffusion model that predicts future thermal baselines.

Input: The model conditions on historical trajectories (observations and actions) and a "held-action" protocol (assuming the previous action is maintained).
Output: It generates a future thermal baseline trajectory ( $\hat{y}_t$ ) representing how the indoor environment would evolve if the current control strategy were held constant. This baseline explicitly captures latent thermal inertia.
Mechanism: The model uses a forward noising process and a learned reverse denoising process to model the distribution of future thermal states, conditioned on the HVAC context.

2. Physics-Aware Regularization

To address the lack of generalization in purely data-driven world models, ADAPT introduces a physics-aware regularization based on a differentiable multi-zone physical heat-balance equation.

Heat-Balance Equation: The generated trajectories are constrained to satisfy a fundamental thermal equation:
$C_i \frac{\hat{T}_{i,k+1} - \hat{T}_{i,k}}{\Delta t} = \hat{Q}_{zone} + \hat{Q}_{out} + \hat{Q}_{rad} + \hat{Q}_{int} + \hat{Q}_{hvac} + d_i$
This equation models inter-zone exchange, outdoor-envelope exchange, solar radiation, internal gains, and HVAC effects.
Joint Training: The model jointly optimizes the diffusion objective and the physics residual. Crucially, the physical coefficients (e.g., thermal conductance) are learned from ground-truth data and then used to regularize the diffusion-generated trajectories. This ensures the model learns invariant thermal mechanisms rather than spurious statistical correlations specific to the training season or climate.
No Manual Calibration: The framework does not require pre-calibrated building parameters; the physics coefficients are learned end-to-end.

3. Integration with Reinforcement Learning

The learned IEWM is integrated into a downstream RL controller to improve decision-making:

Action Evaluation: Before selecting an action, the IEWM predicts a thermal baseline to estimate the delayed impact of maintaining the current action.
Delayed Credit Assignment: After an action is selected, the IEWM performs a "committed-action rollout," imagining future transitions where the selected action is continuously applied. This allows the RL agent to assign credit to actions based on their long-term thermal consequences, rather than just immediate rewards.
Architecture: The framework utilizes a branch-wise action-value network where the thermal baseline is injected into the advantage stream, enabling action preferences to be evaluated with explicit knowledge of future dynamics.

Key Contributions

ADAPT Framework: The introduction of a physics-aware conditional diffusion IEWM specifically designed for HVAC control, capable of predicting delayed thermal dynamics.
Physics-Informed Regularization: A novel regularization technique based on a multi-zone heat-balance equation that enforces physical consistency in diffusion-generated trajectories without requiring manual parameter calibration.
Performance Gains: Demonstrated improvements in both in-distribution (ID) and out-of-distribution (OOD) settings, reducing HVAC energy consumption and occupant discomfort while maintaining robust control across seasonal and climate-region transfers.
Transferability: The IEWM learns transferable thermal dynamics, allowing a controller trained in one season or climate to perform effectively in another without retraining the policy.

Experimental Results

Experiments were conducted on SemiBuildingSim (a high-fidelity simulator based on a real 7-zone office building) and Sinergym (EnergyPlus-based benchmark with cross-climate transfer between Stockholm and Arizona).

In-Distribution Performance (SemiBuildingSim): Compared to the strongest baselines (including MPC, PPO, DreamerV3, and Transformer-based RL), ADAPT reduced HVAC energy consumption by 7.3% and occupant discomfort (PPD) by 30.2%. It also achieved faster convergence and higher sample efficiency.
Out-of-Distribution Prediction: Under seasonal transfer (Summer $\leftrightarrow$ Winter), the physics-aware ADAPT significantly outperformed vanilla data-driven models (e.g., MambaFormer, VAE) in predicting room and return temperatures. The CVRMSE for room temperature was roughly 3-4 times lower than non-physics baselines in OOD scenarios.
Cross-Climate Transfer: In transfers between Stockholm (cold continental) and Arizona (hot desert), ADAPT maintained robust control performance with only marginal degradation relative to in-distribution performance, whereas sequential and data-driven world models suffered substantial performance drops.
Ablation Studies: Removing the physics regularization term or the bias penalty consistently degraded OOD prediction and control performance, confirming the necessity of the physics-aware constraints for transferability.

Significance and Claims

The paper claims that ADAPT offers a practical, energy-efficient, and transferable AI solution for building control science. By explicitly modeling delayed thermal dynamics through a generative world model constrained by physical laws, ADAPT bridges the gap between data-driven flexibility and physical consistency.

The authors assert that this approach enables:

Robust Control: Effective HVAC control that remains stable across changing weather, seasons, and climate regions without requiring target-domain interaction or fine-tuning.
Decarbonization Support: A pathway to reduce building energy consumption and CO2 emissions, aligning with UN Sustainable Development Goals 11 (Sustainable Cities) and 13 (Climate Action).
Generalizability: A framework that can be applied to various building types, from offices to data centers, by learning the underlying thermal physics rather than memorizing specific environmental correlations.

The work positions physics-aware generative modeling as a critical step toward reliable, AI-driven building energy management that can operate safely and efficiently in real-world, dynamic environments.