A long short-term memory physics-informed neural network model for CO2-based natural ventilation rate estimation

Abstract

The occupant-released CO2 tracer gas approach has been widely used for ventilation rate estimation. This approach is non-invasive, low-cost, and does not interfere with the occupants’ activities. However, the CO2 measurement noise and CO2 generation uncertainties can significantly affect the accuracy of the estimated ventilation rates, while the dynamics of the natural ventilation rates could challenge the stability of the estimators. As commonly applied techniques, the moving average filter and the extended Kalman filter have their own advantages and limitations in addressing these issues. To further address the challenges in the natural ventilation rate estimation, this paper proposes a novel model named “NVR-PINN”, based on the long short-term memory-physics-informed neural network and validates it with a case study. The proposed model combines the strengths of the moving average filter and the extended Kalman filter, demonstrating better practical values. It is capable of handling both CO2 measurement noise and CO2 generation uncertainty, effectively capturing the temporal dynamics of the natural ventilation rate, while processing the entire time series observation with a defined sequence window to yield more stable and consistent estimates. The analysis of the case study also revealed useful evidence for relevant research with regard to the applicability of existing ventilation rate estimation techniques.

This research is part of the R&D project IAQ4EDU, reference no. PID2020-117366RB-I00, funded by MCIN/AEI/10.13039/ 501100011033. This work was supported by the Catalan Agency AGAUR under their research group support program (2021 SGR 00341). The author Sen Miao is funded by the China Scholarship Council (CSC) as a full-time PhD student, reference no. 202208390065.

Peer Reviewed

Postprint (author's final draft)