Low-cost SCADA/HMI with Tiny Machine Learning for Monitoring Indoor CO2 Concentration

Concerns about indoor air pollution are increasing as individuals spend much of their time indoors, with carbon dioxide being a notable concern. The increasing interest in developing low-cost gas sensors for indoor air quality monitoring has led to a surge in air quality data generated by these sensing devices. This growing volume of data creates opportunities for implementing machine learning methods in air quality research. However, a challenge of these indoor sensing devices is their resource-constrained memory and computing capabilities, making deploying machine learning algorithms challenging. This paper explores integration of low-cost Supervisory Control and Data Acquisition (SCADA) with tiny machine learning (TinyML) for effective monitoring of current and prediction of future CO2 concentrations. The trained predictors produce RMSE ranging from approximately 0.5-7 ppm when predicting future CO2 concentrations 1, 15, and 30 minutes ahead. Moreover, the models consistently yield confident R-2 scores, ranging from approximately 0.8 to 0.99.