Operator learning for energy-efficient building ventilation control with computational fluid dynamics simulation of a real-world classroom

TL;DR

Addressing the energy-intensive nature of classroom ventilation, this work develops a neural-operator framework that learns the CFD solution operator for spatiotemporal CO$_2$ dynamics and integrates it into a gradient-based optimization to select ventilation supply rates and vent angles under air-quality constraints. The core method uses an ensemble neural operator transformer (based on GNOT) to predict full airflow and CO$_2$ fields from historical data, enabling real-time control with CFD fidelity. CFD-validated experiments on the BEAR-CFD dataset show substantial energy savings (vs Max, Rule-based, and DL-based surrogates) while maintaining safe indoor air quality, with inference speeds exceeding CFD by ~$2.5\times 10^5$ and an open-source dataset for reproducibility. The approach provides a scalable pathway for CFD-informed, real-time building ventilation control and paves the way for multi-zone extensions and real-world deployments.

Abstract

Energy-efficient ventilation control plays a vital role in reducing building energy consumption while ensuring occupant health and comfort. While Computational Fluid Dynamics (CFD) simulations provide detailed and physically accurate representation of indoor airflow, their high computational cost limits their use in real-time building control. In this work, we present a neural operator learning framework that combines the physical accuracy of CFD with the computational efficiency of machine learning to enable building ventilation control with the high-fidelity fluid dynamics models. Our method jointly optimizes the airflow supply rates and vent angles to reduce energy use and adhere to air quality constraints. We train an ensemble of neural operator transformer models to learn the mapping from building control actions to airflow fields using high-resolution CFD data. This learned neural operator is then embedded in an optimization-based control framework for building ventilation control. Experimental results show that our approach achieves significant energy savings compared to maximum airflow rate control, rule-based control, as well as data-driven control methods using spatially averaged CO2 prediction and deep learning based reduced order model, while consistently maintaining safe indoor air quality. These results highlight the practicality and scalability of our method in maintaining energy efficiency and indoor air quality in real-world buildings.

Figures (9)

• Figure 1: Schematic of our data-driven operator learning framework for energy-efficient ventilation control. Computational fluid dynamics (CFD) simulations are used to model complex 3D airflow and CO$_2$ spatiotemporal dynamics. An ensemble of neural operator transformer models is trained to learn the mapping between ventilation control actions and airflow field evolution. Leveraging high-fidelity simulation data, our approach enables real-time optimization of ventilation strategies to minimize energy consumption while maintaining indoor air quality.
• Figure 2: (a) The picture of the studied room and the geometry of the CFD model: a classroom with a ventilation system including 18 inlet vents and 2 outlet vents on the ceiling. Occupants are modeled as a single rectangular cuboid with a prescribed CO$_2$ mass flux to represent CO$_2$ effects of varying occupancy levels. (b) Visualization of the CFD simulation results of CO$_2$ concentration and airflow velocity fields - one example from the developed CFD dataset.
• Figure 3: Overview of the proposed data-driven operator learning framework for energy-efficient ventilation control. The framework consists of two phases: (1) the learning phase, where neural operator transformers are trained to map past air field data and ventilation control parameters to future air field evolution, and (2) the control phase, where the trained ensemble neural operator is integrated into an optimization framework to solve the ventilation control problem. This approach enables real-time optimization of airflow supply rates and vent angles while maintaining air quality standards.
• Figure 4: Training loss (NLL loss) during training (left) and the $l_2$ error for the training (middle) and test sets (right). The $l_2$ error \ref{['eq:relative_error']} is computed based on the ground truth and the model's mean prediction output.
• Figure 5: Operator Learning: Visualization of the ground truth, corresponding predictions from the ensemble model and Model 3, and the relative errors between the ground truth and predictions at the final time step.