S38 Bias Correction for the AIRPACT-5 Model

Sunday, 6 January 2019
Hall 4 (Phoenix Convention Center - West and North Buildings)
Nicole June, The Pennsylvania State University, University Park, PA; and J. Vaughan, Y. Lee, and B. Lamb

Air quality models are used to inform the public of potential health concerns that result when there is a high concentration of pollutants, such as PM2.5 (particulate matter of 2.5 μm aerodynamic diameter), in the atmosphere. The Air Indicator Report for Public Awareness and Community Tracking v.5 (AIRPACT-5) system runs daily to predict concentrations of various pollutants throughout the Pacific Northwest (http://lar.wsu.edu/airpact/index.html).Through comparison with surface PM2.5 measurements, we know that the forecasts have bias, that varies in magnitude during the year, and peaks during wildfire season mainly due to underprediction of PM2.5, in smoke. This project explores multiple methods to correct model bias. We use forecast results and observations at state and local air quality monitoring sites within the AIRPACT-5 domain for January 1, 2017 to December 31, 2017. These data were limited to sites that measured PM2.5 and reported data for the entire year. We reviewed three post-process bias-correction methods including: a subtraction of a rolling mean of the bias, a multiplicative ratio technique, and the Kalman Filter technique. The Kalman Filter is a recursive technique that predicts the future bias based on recent past bias. Previous studies have used this technique to reduce bias of air quality models, including the Community Multiscale Air Quality (CMAQ) model used in AIRPACT-5 (Djalalova et al., 2015). The post-processing method is applied to 24-hour averaged concentrations of PM2.5; a regulatory standard of 35 µg m-3 is stipulated for 24-hour PM2.5 in the National Ambient Air Quality Standards (NAAQS). In application to the AIRPACT-5 results, we found that all the post-processing methods reduce the bias at monitoring stations, including during the wildfire season. Of the three techniques tested, the Kalman filter generated the most accurate results. The raw forecasts had an annual mean absolute error of 3.43 µg m-3, the Kalman filter bias corrected results had an annual mean absolute error of 1.45 µg m-3, a 58% reduction of mean absolute error for the year of 2017. The correlation coefficient between the forecast and observations improved by 13%, from 0.71 for the raw results to 0.82 for the Kalman filter. Unlike the other two methods, the Kalman filter maintained its ability to decrease the error during wildfire season; reducing the mean absolute error by 67% for August, and 30% for September. The mean absolute error for the raw results reached a maximum value of 11.09 µg m-3 in the month of August. The Kalman filter post-processed results had a mean absolute error of 3.59 µg m-3 for the month of August. These results suggest that the Kalman filter bias-correction method might be usefully interpolated to each grid cell in the domain of the model. The use of a cubic spline to spread the bias correction of results spatially throughout the model domain was also investigated. This interpolation method does well in areas of many observation points, but does less well in areas lacking observation sites and at the boundaries of the domain. Further work is needed to improve and further evaluate the interpolation method. Additional work will also focus on using a rolling window of past error for the Kalman filter technique, to better determine the utility in real-time forecasting.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner