FIELD CALIBRATION OF LOW COST AIR QUALITY MONITORS IN TWO CHINESE CITIES WITH NON-LINEAR REGRESSION AND ARTIFICIAL NEURAL NETWORK TECHNIQUES
Abstract
Exposure to fine particular matter (PMv2.5) is associated with various adverse health
outcomes, including cardiovascular disease, cancer and respiratory related diseases.
Reducing exposure assessment errors for epidemiologic studies, and facilitating new
models of community-engaged research may benefit from improved understanding of the
spatiotemporal distribution of PM with respect to mobile and stationary sources of
particular emissions. In this study, the performance of a low-cost PM monitor based on
the Shinyei PPD42NS PM sensor was tested and calibrated with a beta-attenuation based
reference PMv2.5 monitor (BAM-1020) in two Chinese cities (Nanjing and Chengdu) using
linear, power-law and artificial neural network (ANN) approaches. The first eight months’
data in Nanjing showed that the low-cost monitor can provide reasonably accurate
estimation of hourly PM2.5 under non-condensing conditions (RH<95%). Among all three
calibration methods, the ANN approach shows the highest correlation between the
estimated and BAM-1020 measured hourly PM2.5 (R^2=0.76). PM2.5 estimated from the
power-law equation demonstrates a slightly better agreement (R^2=0.70) with BAM-1020
hourly PM2.5 than linear fit method (R^2=0.68). Approximately 73% of the hourly PM2.5
estimated by the low-cost monitor with the ANN calibration approach is within the low-cost
monitor performance guideline of the Ministry of Environment of China, which is
better than linear and power-law approaches (approximately 64% and 67%, respectively).
The better performance of ANN is mainly due to including temperature and relative
humidity (RH) as input data in addition to the raw sensor output of Low-pulse Occupancy
Ratio (LOR). The performance of the low-cost monitor deteriorates after 5-6 months on
continuous operation in polluted environments. R^2 for the first, second and third four-month
periods based on the ANN approach are 0.80, 0.64 and 0.24, respectively. This
suggests that regular replacement or cleaning of the PM optical sensing unit is needed to
use the low-cost monitor for long-term community monitoring. However, in terms of
monthly average concentrations, the low-cost monitor has a small error of 10%, even after
long operation periods. The consistency of the low-cost sensors is also tested in this study.
Poor correlation in sensor raw readings was found among three collocated low-cost
monitors. This suggests that a screening of the sensors of consistency is needed to ensure
consistent results. Also, the calibration parameters developed using the low-cost monitor
data collected in Nanjing do not lead to good estimations of PM2.5 when sensor data in
Chengdu are used. Variation of the sensor-to-sensor responses or different weather
conditions are possible causes, but the root cause of the problem is still unclear and
requires more investigation.
Citation
Wang, Zhenglu (2018). FIELD CALIBRATION OF LOW COST AIR QUALITY MONITORS IN TWO CHINESE CITIES WITH NON-LINEAR REGRESSION AND ARTIFICIAL NEURAL NETWORK TECHNIQUES. Master's thesis, Texas A & M University. Available electronically from https : / /hdl .handle .net /1969 .1 /173509.