10th International Aerosol Conference September 2 - September 7, 2018 America's Center Convention Complex St. Louis, Missouri, USA
Abstract View
Challenges in Low-Cost Sensor Calibration: A Case Study on Deployment of Sulfur Dioxide Electrochemical Sensors in an Urban Environment
REBECCA TANZER, Carl Malings, R. Subramanian, Albert Presto, Carnegie Mellon University
Abstract Number: 605 Working Group: Low-Cost and Portable Sensors
Abstract Low-cost air quality sensors have the potential to greatly increase the spatial resolution of air quality measurements. Low-cost sensors however exhibit cross-sensitivities to gases other than the gas they are built to measure, and sensor response is often dependent on changes in temperature and humidity. In addition, sensor response (e.g., volts/ppb) and the relative importance of cross-sensitivities may change over time. This study examines the precision and bias of low-cost sensors deployed throughout an urban environment over the course of a year, with specific focus on quantifying SO2 measurement errors near a major point source in a SO2 non-attainment area.
We deployed Real-time Affordable Multi-Pollutant (RAMP) sensor packages, which use electrochemical sensors for measurement of gaseous pollutants (CO, NO2, O3, SO2). Previous work from this group and others has shown the utility of machine-learning based calibration algorithms to account for electrochemical sensor cross-sensitivities. Some of these algorithms, such as Random Forests, are unable to extrapolate beyond the range of the training dataset, which is particularly challenging for pollutants like SO2 that often have very low concentrations.
We compare the performance of multiple SO2 calibration models: (1) laboratory-derived linear regression model and (2) Random Forest model built on co-location of RAMPs with a reference SO2 monitor at an urban background location; (3) MLR, (4) Random Forest, and (5) Neural Network models built on co-location calibration immediately downwind of a major SO2 source near Pittsburgh, PA, USA. We held out a portion of the co-location data downwind of the SO2 source for model testing and validation.
The laboratory regression captures the general trend of SO2 concentrations for the testing data at the near-source site, but has high error (root mean squared error, RMSE = 14ppb). The random forest models built at the urban background site perform poorly (coefficient of variation of mean absolute error >80%), likely because the training data set that they were built off of did not incorporate SO2 concentrations >10 ppb. In contrast, all three models built at the high-SO2 site perform better (mean absolute error, MAE < 4.3ppb) presumably because of a more realistic range of training data.
We tested the spatial transferability of the SO2 calibrations built at the high SO2 site by redeploying the RAMPs to new locations. Performance was poorer after redeployment (MAE = 6.7ppb). This appears to be a consequence of the SO2 calibration models being overfit to the specific SO2 source near the calibration location, and therefore less generalizable than similar machine learning models for CO and NO2, which are more variable in time and space. Results for the neural network calibration model generated at the high-SO2 site demonstrated precision error of 44% and bias error of 38%.
Sensors were deployed and maintained over the course of one year (Spring 2017-Spring 2018). Performance of the electrochemical sensors over the course of the deployment period is to be conducted in order to calculate sensor degradation as a function of time.