10th International Aerosol Conference September 2 - September 7, 2018 America's Center Convention Complex St. Louis, Missouri, USA
Abstract View
Calibration and Long-Term Performance Evaluation of Low-Cost Sensors for Gas and Fine Particulate Mass Monitoring with RAMPs
CARL MALINGS, Rebecca Tanzer, Provat Saha, Aja Ellis, Rose Eilenberg, Aliaksei Hauryliuk, Sriniwasa Prabhu Nehru Kumar, Naomi Zimmerman, Levent Burak Kara, Albert Presto, R. Subramanian, Carnegie Mellon University
Abstract Number: 1386 Working Group: Low-Cost and Portable Sensors
Abstract Assessing urban air quality and its variability at a high resolution in space and time requires a large network of monitoring sites. The cost of implementing such a network can be prohibitive if traditional high-precision but high-cost reference instruments are used. As an alternative, the Real-time Affordable Multi-Pollutant (RAMP) sensor system has been developed at the Center for Atmospheric Particle Studies of Carnegie Mellon University, in collaboration with SenSevere (Pittsburgh, PA). The RAMP uses electrochemical sensors to measure concentrations of up to four gaseous pollutants out of carbon monoxide (CO), sulfur dioxide (SO2), nitric oxide (NO), nitrogen dioxide (NO2), Ozone (O3), and volatile organic compounds (VOCs), and also includes sensors for temperature and relative humidity. Furthermore, it can interface with external sensors measuring fine particulate matter (PM2.5).
Calibration models are developed for each of the electrochemical gas sensors as well as the Met-One Neighborhood Particulate Monitor (NPM) and PurpleAir external PM sensors deployed with the RAMPs. For the electrochemical sensors, cross-sensitivities with other gases as well as temperature and relative humidity impact the responses of the sensors. For the particulate sensors, corrections for aerosol hygroscopic growth at high humidity and local aerosol size distribution relative to sensor detection limits are required.
Various approaches were evaluated to generate calibration models to match raw sensor readings to reference-grade instruments with which the RAMPs were collocated prior to deployment. Data from the collocation are divided into a set used to train models and a separate testing set for validation. Simple linear and quadratic regression were applied, along with multiple nonparametric approaches. A nearest-neighbors clustering algorithm matches new measurements to similar observations in the training data. A neural network uses layers of simple operations to perform complicated nonlinear transformations. A random forest model uses sets of decision rules to group measurements and averages results across these sets. A hybrid random forest model combines random forest and linear models to generalize beyond the training data range. For each approach, separate models were calibrated for each RAMP, and generic models were also created which were applicable across all RAMPs.
Performance of these calibration techniques was assessed in terms of correlation (measured by Pearson correlation coefficient, R) and estimation error (coefficient of variation of the mean absolute error, CvMAE). Simple linear and quadratic models were found to work sufficiently well for CO and O3 sensors, having comparable performance metrics with more sophisticated nonparametric models (median R above 0.9, median CvMAE below 0.2). For other electrochemical sensors, the hybrid random forest model tended to have the best overall performance. However, performance varied among these sensors, with SO2 having a relatively low error (median CvMAE 0.34) but also low correlation (median R 0.45), while for NO2 correlation was relatively better (median R 0.59) but error was worse (median CvMAE 0.47). Compared to RAMP-specific models, general models tended to have comparable performance (average CvMAE increase of 0.1, average R decrease of 0.05).
For the external PM sensors, a correction using RAMP measurements and an average aerosol chemical composition and size for the city of Pittsburgh was applied first. This was followed by additional empirical linear corrections based on collocation with regulatory monitors. Across 12 NPM sensors, median bias was 0.3 µg/m3 (range -1.8 to +5.7 µg/m3), median mean absolute error was 3.4 µg/m3, and median R was 0.93. We shall present results from similar testing of 36 NPM sensor and 12 PurpleAir sensors from such collocations and the performance of the RAMP gas and external PM2.5 sensors during long-term (over 12 months) field deployment.