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

Building energy consumption constitutes roughly one-third of global energy use, with HVAC systems being the dominant contributor. Traditional control approaches like MPC and DRL have achieved some success but face limitations in modeling delayed thermal responses and environment variability. Recent advances in generative models, especially diffusion models, have shown promise in long-horizon prediction tasks, yet their application in building thermal dynamics remains underexplored. Existing models often lack physical interpretability and transferability, leading to poor performance under distribution shifts. Integrating physical laws into deep generative models offers a promising avenue to overcome these challenges, enabling more robust and generalizable indoor environment predictions.

The core challenge is accurately modeling the delayed and partially observable thermal responses of buildings across diverse environments. Existing models tend to overfit to training data, lacking physical consistency, which hampers their transferability. This results in suboptimal HVAC control, increased energy consumption, and occupant discomfort, especially under unseen seasonal or regional conditions. Developing a model that can explicitly capture latent thermal inertia, adhere to physical laws, and generalize across environments is critical for advancing energy-efficient, occupant-centric HVAC systems.

This work introduces a physics-aware conditional diffusion model that explicitly predicts future thermal baselines, capturing latent thermal inertia. The key innovations include:

• �� Using a diffusion probabilistic framework for flexible, high-fidelity trajectory generation.
• �� Incorporating a differentiable multi-zone heat-balance regularizer to enforce physical consistency without manual parameter calibration.
• �� Enabling transfer across seasons and climate zones, addressing the limitations of purely data-driven models.
• �� Supporting downstream reinforcement learning with explicit delayed thermal predictions, improving control robustness and efficiency.

These innovations collectively enable a physically grounded, transferable, and robust indoor environment model, significantly advancing the state-of-the-art in building energy management.

• �� Model formulation as a Partially Observable Markov Decision Process (POMDP), with partial sensor observations including zone temperatures, humidity, occupancy, outdoor weather, and solar radiation.
• �� Design of a conditional diffusion process: input historical trajectories and future control sequences, generate future thermal baselines via a reverse denoising process.
• �� Implementation of a multi-zone heat-balance regularizer, modeling inter-zone heat transfer, outdoor exchange, solar gains, internal heat, and HVAC effects.
• �� Joint training of the diffusion model with the physical regularizer, optimizing a combined loss function that includes the diffusion denoising objective, physical residuals, and parameter identification.
• �� Integration into a reinforcement learning framework: the predicted thermal baseline supports delayed reward attribution, while imagined future rollouts improve policy learning.
• �� Use of branch-wise action-value networks, with the thermal baseline injected into advantage streams for more informed decision-making.

Experiments utilize SemiBuildingSim, based on real building data, and Sinergym, simulating different climates. Metrics include CVRMSE, occupant comfort (PPD, PMV), energy consumption, and control smoothness. The models are trained with hyperparameters tuned for diffusion steps and regularization weights. Transfer tests involve switching between seasons and climates, assessing robustness. Baselines include MPC, traditional RL, and pure data-driven models. The evaluation emphasizes prediction accuracy, control effectiveness, and transfer robustness, with ablation studies validating the importance of physics regularization.

ADAPT outperforms baselines with a 7.3% energy reduction and 30.2% occupant comfort improvement in IID settings. In cross-season and cross-region tests, it maintains high performance with minimal degradation, significantly better than non-physics models. The physics regularizer enhances physical plausibility, reduces prediction errors, and improves robustness against environment shifts. These results demonstrate the model’s capacity for reliable, transferable indoor environment prediction and control.

This approach can be integrated into smart building management systems for energy savings and occupant comfort. It is suitable for existing buildings with limited sensor data, as it requires only partial observations. The model’s transferability makes it ideal for multi-region deployment, supporting urban sustainability goals. Future deployment will involve sensor integration, edge computing, and real-time inference to enable adaptive, energy-efficient HVAC control.

The training process demands extensive high-quality data, which may be costly. Diffusion inference involves iterative steps, increasing computational load, challenging real-time applications. Although physics regularization improves robustness, extreme climates or unique building geometries may still pose challenges, requiring further physical parameter calibration. Future work should focus on reducing inference costs and enhancing adaptability to diverse physical structures.