Experimental validation of neural network-based prediction of natural ventilation bulk airflow rate

• New soft sensor for Natural Ventilation bulk airflow rate. • Soft sensor demonstrated good accuracy in the experimental validation, with average error of 27%. • A Neural Network model was trained using soft sensor data to predict NV flow rates. • The Neural Network prediction model demonstrated moderate accuracy, with average error of 30%. To fully exploit natural ventilation (NV) as an energy-saving strategy in mixed-mode buildings, accurate real-time prediction of NV airflow rates is essential. Current approaches for NV airflow rates prediction often rely heavily on expertise knowledge and computationally demanding methods such as Computational Fluid Dynamics (CFD) as well as expensive and complex direct airflow measurements. These methods pose significant practical limitations, hindering widespread implementation in building management systems (BMS). To address this research gap, this study proposes and experimentally validates a novel methodology to support advanced real-time NV control strategies. This methodology is cost-effective, scalable, and easily deployable, eliminating the need for expensive equipment or extensive modeling expertise. To meet the challenge of creating a real time predictor of NV airflow this paper proposes a two-stage approach. In the first stage a soft sensor to measure bulk airflow is developed and validated using experimental data obtained in a large mixed mode academic building. In the second stage, the prediction of the soft sensor is used to train a data-driven neural network (ANN) model for NV flow rate prediction. The soft sensor relies on thermal zone sensible heat balance to measure NV flow rates using commonly available building management system (BMS) data. The average estimation error of the soft sensor, when compared to CO 2 decay based measurements, was 27%. The estimations of the soft sensor were used to train a data-driven artificial neural network (ANN) model to predict NV airflow rates. The ANN was structured as a multi-layer perceptron (MLP), using key input features to characterize the wind and buoyancy-driven forces, including indoor and outdoor air temperatures, indoor-outdoor temperature differences, effective window opening areas, and wind speed and direction. The ANN model was trained on data collected over several months, encompassing diverse environmental conditions and multiple NV configurations. The ANN’s predictive accuracy was validated by comparing it to CO 2 decay-based measurements, achieving a mean absolute percentage error (MAPE) of approximately 30%. The proposed model has autonomous self-calibration capabilities and does not require CFD simulations. It is simple to deploy and does not require expert intervention. In the current dataset the model is able to predict flows that range from 0.6-6 m 3 /s (one order of magnitude variation) with moderate/good accuracy. By eliminating the need for expensive direct airflow measurements or computationally intensive CFD simulations, this methodology holds substantial potential for broad implementation in real-world smart building applications. The scalable and adaptable nature of the proposed approach ensures its seamless integration into existing BMS setups, enabling optimized and automated control of NV systems, reducing reliance on mechanical cooling, and enhancing overall building energy efficiency.