Current Issue

A three-dimensional air quality forecasting system consisting of Weather Research and Forecast (WRF) model v3.4.1 (Skamarock and Klemp, 2008), Sparse Matrix Operator Kernel Emissions Processor (SMOKE) v2.1 (http://www.smoke-model.org), and the Community Multiscale Air Quality (CMAQ) modeling system v4.7.1 (Byun and Schere, 2006) was used for PM_2.5 simulation. WRF model simulations were initialized with Global Forecasting System (GFS) data sets. The WRF model results were prepared for daily emission processing and air quality simulations using the Meteorology–Chemistry Interface Processor. The Statewide Air Pollution Research Center version 99 (SAPRC99) and the fifth-generation modal aerosol model (AERO5) were used as the chemical mechanism and aerosol module, respectively, for the CMAQ simulation.

For anthropogenic emissions, the Intercontinental Chemical Transport Experiment-Phase B (INTEX-B) inventory for the year 2006 (Li et al., 2014; Zhang et al., 2009) was used for Northeast Asia, and the Clean Air Policy Support System (CAPSS) inventory for the year 2007 was used for Korea (Lee et al., 2011; Kim et al., 2008). Biogenic emissions were obtained using the Model of Emissions of Gases and Aerosols from Nature (MEGAN) version 2.04 (Guenther et al., 2006). Fig. 1 shows the modeling domain consisting of two grids with horizontal resolutions of 27 and 9 km. There were 15 layers vertically on a sigma coordinate up to 50 kPa with the lowest layer thickness of about 32 m. Default profiles provided with CMAQ were used as the boundary conditions for the coarse grid, and the boundary conditions for the fine grid were updated by the model outputs from the coarse grid.

PM_2.5 samples were collected on a Teflon filter (Zefluor, Pall) using a well impactor ninety-six (WINS) and a sequential sampler (PMS-103, APM) at a flow rate of 16.7 L/min for 24 hours. Concentrations of PM_2.5 and inorganic ions were measured using an automated filter weighing system (MTL) equipped with a microbalance (UMX2, Mettler Toledo) and ion chromatography (ICS 2000, Dionex), respectively. PM_2.5 samples were also collected on a quartz filter (Tissuquartz 2500QAT-UP, Pall) to measure concentrations of organic and elemental carbons using an OCEC analyzer (Sunset). Concentrations of PM_2.5 and its components were available on 460 days (47%) at Seoul (SL), 410 days (42%) at Chungcheong (CC), 456 days (47%) at Honam (HN), 329 days (34%) at Yeongnam (YN), and 288 days (30%) at Gangwon (GW), out of 975 days during the study period.

The model performance was evaluated using mean fractional bias (MFB), correlation coefficient (R), and the slope and interceptor of best-fitted line between predicted and measured values. MFB is defined by

where p_i and m_i denote predicted and measured values, respectively, and N denotes the number of data (Boylan and Russell, 2006). We adopted the performance goals and criteria suggested by Boylan and Russell (2006), which denote the levels of accuracy that the best model can achieve and that are acceptable for standard model applications, respectively. They were given as:

where C¯ is (p¯+m¯)/2 in μg/m³, and p¯ and m¯ are the means of predicted and measured values, respectively.

During the study period, PM_2.5 concentrations were forecasted by dividing them into four levels as follows: good (≤15 μg/m³), moderate (15-50 μg/m³), bad (50-100 μg/m³), serious (>100 μg/m³). The forecasting performance was evaluated by examining whether the predicted level agreed with the measured level. We distinguished four groups from “A” to “D” and another four groups from “e” to “h” (Fig. 2). “A” indicates that low measured concentrations, which fall into either the good or moderate level, were predicted as high concentrations, which fall into either the bad or serious level. “B” indicates that high measured concentrations were correctly predicted. “e” to “h” indicates that each level was correctly predicted. Because both predicted and measured concentrations were divided into either high or low concentrations (in “A” to “D”), the sum of “A” to “D” is 100%.

Fig. 2. 
Definition of parameters for forecasting performance statistics.

We defined four parameters—the accuracy, probability of detection (POD), false alarm rate (FAR), and bias ratio—as shown in Fig. 2 (NIER, 2014; McKeen et al., 2005; USEPA, 2003). The accuracy is the percent of forecasts that correctly predicted the concentration levels. The remaining three parameters examine the quality of high-concentration forecasts. POD represents the ability to correctly predict high-concentration events, whereas FAR is the percent of high-concentration predictions that did not occur. The bias ratio is the ratio of predicted high-concentration events to observed high-concentration events. A bias ratio greater than 1 indicates that high-concentration events are overpredicted.

Fig. 3 compares the major components between measured and predicted concentrations by station. The measured PM_2.5 concentration is lowest at YN and highest at GW. YN exhibits the lowest concentrations of major components except for the unresolved other component whose concentration is highest. At GW, the OC concentration is highest whereas the secondary ions are lowest except for YN. Underestimation of the predicted PM_2.5 is remarkable at GW (Table 1), where the measured PM_2.5 is highest. In Table 1, the ratio of predicted to measured concentration for NH₄⁺ is close to one on the whole. However, the ratio for NO₃^- is greater than 1, and the ratios for carbonaceous components (OC and EC) are less than 0.3, indicating significant underestimation. For the major ions, the overestimation of NO₃^- at YN and the underestimation of SO₄^2- at GW are notable. At GW, OC underestimation is serious, and EC is underestimated similarly to CC and HN. Despite a substantial overestimation of NO₃^- at YN, its effect on PM_2.5 is insignificant because of a low proportion (Fig. 3). Given a prevalence of underestimations, the model performance is better at YN and SL because of high ratios of predicted to measured concentration for major components.

Fig. 3. 
The mean values of measured and predicted concentrations of major components at monitoring stations by region. The sum of the components is PM_2.5, and the other is the remainder of PM_2.5 excluding the components shown in the figure.

Table 2 shows the differences in model performance metrics by station due to bias correction. The predicted concentrations become identical to the measured concentrations because the bias was corrected by multiplying the predicted concentration by the ratio of measured to predicted concentration (ratio adjustment). In contrast, R and the relative intercept remain unchanged. On the whole, the model performance was improved by the bias correction, as MFB moves within the goals from outside and the slope increases from 0.62 to 0.95. The effect of bias correction is noticeable at GW, where MFB moves within the goals from outside the criteria and the slope increases from the lowest at 0.47 to 0.89. At HN, MFB falls within the goals after correction, but its absolute value is still highest along with that at CC, and the slope increases above 1.0, indicating that the correction effect is unclear.

Table 3 shows the differences in MFB for major components. Overall, MFBs for SO₄^2- and the other, which fall outside the criteria and goals, respectively, move within the goals. In contrast, MFB for NO₃^- is pushed outside the criteria because of the correction. MFBs for both OC and EC are improved, but are still outside the criteria. Looking into the differences by station, SO₄^2- improves at all stations, as does the other at all stations except for YN where MFB originally fell within the goals. On the other hand, NO₃^- and NH₄⁺ exhibit degradation at all stations; particularly, MFBs for NO₃^- at SL, CC and HN move outside the criteria, and MFB for NH₄⁺ at GW moves outside the goals.

Fig. 4 shows a plot of predicted vs. measured concentrations for PM_2.5 at all stations. Individual values are compared, different from comparing mean values in the previous sections to examine the model performance. Originally, more data points lie below the 1 : 1 line, indicating the tendency of underestimation of predicted concentrations (Fig. 4(a)). However, the data points move upward due to bias correction, and the amount of data whose predicted level coincides with the measured level increases, despite some overpredicted data points.

Fig. 4. 
Plot of predicted vs. measured PM_2.5 concentrations at all stations. Dotted lines denote the division of concentration levels, and solid lines denote the division of high and low concentrations (see Fig. 2 and the description in the text for details). “Correct” and “false” in the legend indicate that the predicted level coincides and does not coincide with the measured level, respectively. The biases were corrected using mean values by station in (b).

The differences in the forecasting performance statistics that resulted from bias correction are summarized in Table 4. Originally, the overall accuracy for all levels was moderate at 58%, but POD for high-concentration events was only 23% (Table 4(a)). FAR and the bias ratio are also low at 33% and 34%, respectively. High-concentration forecasts are generally fewer, particularly at GW, and consequently, high FAR and the bias ratio at YN are distinguished. Table 4(b) shows the bias corrected performances. Overall, the accuracy and POD increase by 10% and 30%, respectively, whereas FAR also increases to 56%. Most of all, because the frequency of high-concentration forecasts greatly increases, the bias ratios exceed 100% except for SL. By station, all four parameters at HN and GW greatly increase. The differences in performances between the stations are generally reduced by the bias correction, although a high value of POD at HN becomes even higher. A representative case is GW where FAR increases from 0% to 49%.

Note that mean values during the same period were used for bias correction, which cannot be accomplished in the real-time forecasting. However, we tested this bias correction because the effectiveness of bias correction using mean values did not depend much on the period of data used for the correction in our previous study (Ghim et al., 2017). It was probably because the biases of model results from measurement data in Korea were systematically caused by limitations in reproducing the atmospheric environment such as meteorology and emissions during model simulation. The present study revealed that the biases were specific to station (or region) and that the correction should be conducted by station (or region).