Low-Cost CO2 Sensors: On-Site Performance Evaluation and Co-Location Correction Procedure for Reliable Ventilation Assessments in Schools

TL;DR

This paper evaluates low-cost CO2 sensors for school ventilation monitoring and introduces a method to improve their accuracy without lab equipment.

Contribution

A practical on-site calibration method for low-cost CO2 sensors using co-location and normalization is introduced.

Findings

Regression-based correction reduced sensor error by 16% during validation.

Final correction factors reduced RMSE by 27% and out-of-range measurements by 43%.

Corrected sensors showed RMSE of 7.4 ppm at ambient CO2 levels and 11.9 ppm below 1500 ppm.

Abstract

Adequate ventilation is essential for maintaining indoor environmental quality in schools, where ventilation standards are often based on an indoor concentration of human-generated carbon dioxide (CO2) above ambient levels. Low-cost non-dispersive infrared (NDIR) CO2 sensors offer a practical solution for ventilation monitoring, yet variability between sensors can compromise accuracy, particularly when applications depend on the determination of precise concentration differences. This study evaluates the performance of twenty-three low-cost CO2 sensors, developing normalisation functions to improve comparability across sensors, introducing an accessible methodology for on-site sensor calibration without the need for laboratory-grade reference equipment. The sensors were co-located for three independent test periods in 2025 representing typical school internal conditions in Ireland. Pre-normalisation analysis showed strong linearity (coefficient of determination (R2) = 0.999) but notable variability, with a mean root mean square error (RMSE) of 18.3 ppm and 0.45% of measurements outside manufacturers stated accuracy. Normalisation models were trained and validated using a leave-one-period-out approach. Regression-based correction yielded the greatest improvement, reducing RMSE by 16%. When applied to the full dataset, final correction factors reduced RMSE by 27%, out-of-range measurements by 43%, and proportional bias by 31%. Corrected sensors demonstrated highly consistent performance, particularly within the CO2 ranges most relevant for classroom ventilation assessment, with an RMSE = 7.4 parts per million (ppm) at ambient concentrations and 11.9 ppm at concentrations below 1500 ppm. Field-based co-location in the deployment environment across full CO2 cycles, combined with a network-derived global reference, produced effective correction factors. Performance declined marginally above 1500 ppm and during dynamic occupancy, while overall accuracy remained strong. The study presents a practical and accessible methodology for evaluating and normalising low-cost CO2 sensors without specialised laboratory equipment, supporting reliable ventilation assessments in schools.

Figures