AI Building EnergyEvidence Atlas
Home / Topics
Research theme

Physics-Informed Building Energy Modeling

Four routes are represented: grey-box dynamics coupled with neural disturbance forecasting, physical neurons embedded in network architecture, physically consistent neural networks used in reinforcement learning, and RC-informed neural ordinary differential equations for small datasets. The common objective is to improve generalization, interpretability, robustness, or control reliability beyond purely data-driven models.

Comparison between black-box neural networks and EP-PINNs containing physical neurons.
Figure 3. Network-architecture comparison between conventional black-box models and EP-PINNs with embedded physical neurons.
Figure excerpt from Fang et al. (2025), Building Simulation. DOI 10.1007/s12273-025-1293-z
Source-attributed paper figure
Physically consistent neural network integrated with deep reinforcement learning for residential heating control.
Figure 1. Proposed physically consistent neural network and deep-reinforcement-learning control framework for residential space heating.
Figure excerpt from Ran et al. (2025), Building Simulation. DOI 10.1007/s12273-025-1253-7
Source-attributed paper figure
Seasonal naive neural ordinary differential equation workflow for building load prediction.
Figure 1. Overall workflow of SN-NODE for building-load prediction, robustness analysis, and multi-resolution forecasting.
Figure excerpt from Ma et al. (2025), Building Simulation. DOI 10.1007/s12273-025-1269-z
Source-attributed paper figure

Questions addressed

  • What are the main forms of physics-informed building energy modeling?
  • How can physical structure improve extrapolation and small-data performance?
  • Which physics-informed models have been tested in HVAC control?
Source figures

Methods and applications represented in this theme

Comparison between black-box neural networks and EP-PINNs containing physical neurons.
Figure 3. Network-architecture comparison between conventional black-box models and EP-PINNs with embedded physical neurons.
Figure excerpt from Fang et al. (2025), Building Simulation. DOI 10.1007/s12273-025-1293-z
Source-attributed paper figure
Physically consistent neural network integrated with deep reinforcement learning for residential heating control.
Figure 1. Proposed physically consistent neural network and deep-reinforcement-learning control framework for residential space heating.
Figure excerpt from Ran et al. (2025), Building Simulation. DOI 10.1007/s12273-025-1253-7
Source-attributed paper figure
Seasonal naive neural ordinary differential equation workflow for building load prediction.
Figure 1. Overall workflow of SN-NODE for building-load prediction, robustness analysis, and multi-resolution forecasting.
Figure excerpt from Ma et al. (2025), Building Simulation. DOI 10.1007/s12273-025-1269-z
Source-attributed paper figure

Comparison of selected papers

PaperResearch problemMethodReported evidence
Yang et al. (2025)High-inertia thermally activated building systems are difficult to control when key disturbances such as occupancy and solar irradiance are stochastic or unmeasured.A hybrid MPC framework combines grey-box system modeling with neural-network disturbance forecasting and compares no-, single-, and multi-disturbance prediction strategies.Relative to conventional MPC, disturbance-aware MPC reduces temperature violations by 20%–42%, cuts cost by 6%, and improves energy flexibility by 3.1%–8.6%.
Fang et al. (2025)Chiller models trained on sparse operating data may extrapolate poorly and provide limited physical interpretability.EP-PINNs embed physical neurons directly in the network architecture and use simplified physical-loss terms during training.The method is tested on four chillers and is reported to maintain high accuracy while improving generalization and interpretability relative to purely data-driven models.
Ran et al. (2025)Residential heating control under cold snaps requires models that are accurate, robust, interpretable, and efficient enough for control optimization.A physically consistent neural network is compared with RC and conventional data-driven models, then embedded in a deep-reinforcement-learning controller.The abstract reports up to 93.9% lower MAE than the RC model, mean indoor temperature of 21.9°C, 69.2% PV consumption, about 37% lower grid imports, and nearly 48% lower energy cost.
Ma et al. (2025)Conventional machine-learning energy models can degrade sharply when only small training datasets are available.The seasonal naïve neural ODE model incorporates an RC-informed structure and introduces a nonlinear robustness index.The reported hourly and sub-hourly CVRMSE values are about 0.3 and 0.2; the model remains effective with only two weeks of training data and outperforms LSTM under comparable data scarcity.

Paper pages