building energy simulation

Existing building thermal datasets lack sufficient diversity in building types and operational conditions, limiting advances in thermal dynamic modeling, energy efficiency control, and fault diagnosis. This work addresses this gap by constructing and releasing the ThermBuild dataset, which uniquely integrates 15 months of high-resolution (15-minute) measured data from two real residential buildings with three years of TRNSYS-simulated data from 958 residential units. The dataset encompasses diverse heat pump systems, building characteristics, and climatic conditions, and includes multidimensional variables such as heat pump operation, indoor environmental parameters, and weather data. ThermBuild enables research in cross-domain transfer learning, simulation-to-reality generalization, and model benchmarking, significantly enhancing the applicability, robustness, and reproducibility of data-driven approaches in building energy systems.

High-quality, large-scale empirical data are scarce in machine learning research on building thermal dynamics. Method: This paper proposes a low-barrier, scalable synthetic data generation framework that integrates a Modelica-based single-zone thermal model with Functional Mock-up Unit (FMU) export capabilities, enabling fully automated Python-driven simulations without requiring domain expertise in building simulation. Contribution/Results: The framework significantly enhances data scale and configuration flexibility compared to existing tools. It generates a large-scale dataset suitable for transfer learning and validates fine-tuning across 486 data-driven models. Experimental results demonstrate superior effectiveness, generalizability, and scalability, establishing a robust data infrastructure for AI-driven building research.

Physics-based models in building energy modeling offer high interpretability but suffer from poor generalizability, while data-driven models achieve high accuracy only with large-scale, high-quality datasets. Method: This paper proposes a hybrid modeling framework integrating mechanistic and data-driven approaches. It introduces a novel dynamic-weight coupling mechanism and physics-constrained training strategy, synergistically combining thermodynamic equations, graph neural networks (GNNs), and physics-informed neural networks (PINNs). End-to-end optimization is achieved via a joint loss function and a differentiable building energy simulator. Contributions/Results: Evaluated on five real-world building datasets, the framework achieves a mean energy consumption prediction error of only 2.3%, reducing error by 47% compared to purely data-driven methods. It also attains a threefold improvement in inference speed. The approach significantly enhances fidelity and robustness under low-data regimes and diverse operational conditions, enabling real-time energy efficiency diagnostics and control.

Existing hybrid approaches for probabilistic building energy modeling overlook inherent uncertainties from weather and occupant behavior and lack systematic probabilistic benchmarking. Method: We propose a physics–data dual-driven probabilistic modeling framework, establishing a unified probabilistic benchmark covering five hybrid paradigms; employing feedforward neural networks for residual learning—using physics-based model outputs as inputs or constraints—and integrating quantile conformal prediction for uncertainty calibration. Contributions/Results: We first reveal the universal superiority of residual learning in quantile forecasting; empirically validate quantile conformal prediction’s effectiveness in calibrating indoor temperature prediction intervals; and demonstrate that our framework achieves both high predictive accuracy and physical plausibility across diverse room types and out-of-distribution scenarios, significantly enhancing robustness and calibration reliability.

This work addresses the lack of a generalizable thermal dynamics model in building energy systems that can transfer across diverse buildings, climates, and control strategies. The authors propose a decoder-only Transformer architecture infused with physical priors, embedding thermodynamic laws through derivative augmentation and Euler-based numerical integration. By integrating static building characteristics and rotary positional encoding (RoPE), the model achieves universal representation capabilities. Notably, this is the first approach to incorporate physics-informed constraints directly into a Transformer framework, yielding a foundational thermal model capable of zero-shot transfer. Evaluated on the CityLearn dataset, the model achieves single-step prediction RMSE as low as 0.29–0.30°C after training on only two buildings, and demonstrates strong generalization to unseen buildings and climate zones, significantly outperforming both conventional methods and fine-tuned temporal foundation models.

Existing building thermal dynamics datasets are predominantly derived from steady-state operation under fixed control policies, which inadequately excite the system’s state space and consequently limit the generalization capability of data-driven models. To address this, this work proposes BuilDyn, a novel toolkit that introduces, for the first time, a control-oriented active excitation mechanism. By combining sampling across building parameter distributions with customizable excitation strategies, BuilDyn generates thermal dynamics data with high state-space coverage and integrates seamlessly into machine learning workflows via a Python API. Experimental results demonstrate that models trained on BuilDyn-generated data significantly outperform those based on conventional datasets in both predictive accuracy and generalization, thereby establishing a robust data foundation for control-oriented modeling, transfer learning, and the development of building-specific foundation models.

Traditional building energy simulation is computationally expensive, and existing surrogate models struggle to generalize across regions, particularly underperforming in data-scarce new locations. This work proposes a week-scale weather-guided surrogate modeling approach that captures short-term weather–energy consumption dynamics common across multiple sites, enabling the construction of a universal model capable of cross-climate-zone transfer without requiring multi-site training. By integrating high-resolution weather feature embeddings, cross-regional pattern alignment, and a data-efficient deep learning architecture, the method achieves near-lossless cross-location prediction within a single climate zone and only modest performance degradation across distinct climate zones. This significantly enhances the model’s scalability, reusability, and generalization capability.

This work addresses the limitation of current large language models (LLMs) in multi-zone HVAC control, which typically lack explicit modeling of building physics and thermodynamic processes. To bridge this gap, the authors propose a knowledge graph that integrates thermodynamic principles with spatial semantics, constructed upon the Brick ontology and enriched with historical environment-controller interaction data to provide LLMs with structured contextual information. This approach represents the first integration of physics-informed spatial semantic graphs into an LLM-based control framework, explicitly capturing inter-zone thermal couplings and building dynamic responses. Evaluated in a five-zone building simulation, the method significantly improves the trade-off between energy efficiency and occupant comfort compared to both conventional and existing LLM-based strategies, achieving the lowest PMV violation rate while maintaining high energy performance.

This work addresses the high computational cost of high-fidelity simulation for large-scale HVAC systems by proposing a scalable, data-driven framework that integrates physics-informed neural ordinary differential equations (PINODEs) with differential-algebraic equation (DAE) solvers. At the component level, implicit PINODEs model heat exchanger dynamics while preserving key physical conservation laws; at the system level, a DAE solver enforces nodal constraints, complemented by a lightweight correction network to mitigate error accumulation. The approach further incorporates gradient-stable latent variable evolution, a neural–DAE coupled architecture, and Bayesian optimization for adaptive solver parameter tuning. Evaluated on both dual-compressor and 32-compressor–condenser systems, the method achieves several-fold acceleration over conventional solvers while maintaining high accuracy, with mean absolute percentage errors (MAPE) consistently below a few percent, thereby offering an effective balance between precision and scalability.