Optimization and Evaluation of Calibration for Low-cost Air Quality Sensors: Supervised and Unsupervised Machine Learning Models

Abstract

With the advancement of air pollution management, low-cost sensors are increasingly being used in air quality monitoring, but the data quality of these sensors is still a major source of concern. In this paper, data from five air monitoring stations in Sofia were compared to data from fixed low-cost PM sensors. The values of atmospheric pressure from low-cost sensors and the effects of relative humidity were investigated. A two-step model was created to refine the calibration process for low-cost PM sensors. At first, we calibrated the sensors with five separate supervised machine learning models and then the ANN-final model with anomaly detection completed the results. The ANN-final model improved the R2 values of the PM10 determined by low-cost sensors from 0.62 to 0.95 as compared to standard instruments. In conclusion, the two-step calibration model proved to be a positive solution to addressing low-cost sensor efficiency issues.