Satellite measurements oversee China's sulfur dioxide emission reductions from coal-fired power plants

The bottom-up SO₂ emissions from coal-fired power plants for the period 2005–2012 were used for comparison with the OMI-derived SO₂ emission burdens. These emissions data are available from a unit-based power plant emission inventory for Mainland China that updated the dataset in Wang et al [25] We calculated the monthly emissions for each unit based on its technology and operation information, including boiler size, coal consumption per unit electricity supply, emission control technology, and the exact month in which the unit formally came into operation and closed. In the bottom-up emission inventory, all FGD facilities were considered as full load operations beginning from their formal installation date. The SO₂ removal efficiencies of FGD facilities were taken from official unpublished data offered by the Ministry of Environmental Protection (MEP) of China.

OMI onboard NASA's EOS-Aura satellite is a nadir viewing, UV/VIS solar backscattering spectrometer, which provides simultaneous retrievals of atmospheric tracer gas and aerosol concentrations, including SO₂, with nearly daily global coverage and a local afternoon equator crossing time at 13:45 since October 2004 [26]. In this study we began with the operational OMI Level-2 planetary boundary layer (PBL) SO₂ product publicly released by NASA [27].

In the work presented here, a number of efforts toward improving the original SO₂ columns have been initiated following the method of Lee et al [28]. Two noteworthy aspects are as follows: (1) daily removal of systematic offsets and substantial latitude dependent biases using the 'reference sector method' and (2) recalculation of the coincident local air mass factors (AMFs) using a radiative transfer model (LIDORT) [29] with a priori SO₂ profiles at a resolution of 2° × 2.5° from a global chemical transport model (GEOS-Chem) [30] for all observations. The retrieved vertical SO₂ columns have been validated against airborne in situ measurements, showing significant improvement in accuracy compared to the original operational data product [28]. Only OMI pixels with solar zenith angle ≤70°, surface albedo ≤0.3 and cloud radiance fraction ≤0.3 were used in the final gridded columns. The outermost one-third of the pixels was excluded to limit the original pixel sizes from 13 × 24 km² at nadir to approximately 16 × 40 km² at the edges. Pixels specified as 'row anomalies' since June 2007 were also dynamically removed, according to the recommendations of NASA. Finally, all pixels were allocated by area weighted mean approach into 0.01° × 0.01° grids with corner coordinate information, in which each pixel over the target grid was weighted by its overlapping area to generate the average SO₂ columns at high-resolution [31].

We first selected isolated power plants by the proportion of the power plant SO₂ emissions in the total anthropogenic emissions (>80%) in the corresponding 0.1° × 0.1° grid and total population (<0.5 million). We then used the high-resolution OMI measurements to estimate the SO₂ emission burdens around 26 isolated coal-fired power plants which passed the 2D Gaussian fitting. Detailed information of these power plants is presented in table S1 and their locations are denoted in figure 1(a). Fioletov et al [19] reported that the elliptically distributed SO₂ columns observed by OMI over a long time period nearby the 40 largest US emission sources could be well fitted by a 2D Gaussian function and the total number of SO₂ molecules, namely the SO₂ burden, associated with the target source could be inferred from the fitted parameters. Results in their study suggested that the OMI-derived SO₂ burdens were correlated well with the emissions directly measured by the continuous emissions monitoring systems, indicating this inferred quantity as a reliable linear indicator of the associated source strength. de Foy et al [23] further pointed out that the fitted emissions (derived with an assumed lifetime) with this method were sensitive to the domain choice. The three-year averaged OMI SO₂ columns (figure 1(a)) at a resolution of 0.01° × 0.01° based on area-weighted method also suggest clearly enhanced emission signals near many coal-fired power plants selected in this study. However, more challenging in China than in US are the various interacting emission sources. In China, coal-fired power plants are typically collocated with populous clusters, which include numerous industrial and residential emission sources and release even more SO₂ than a single power plant. Although we limited our targets only to those far away from megacities and densely polluted regions, perfectly isolated coal-fired power plants are scarce and only account for 6 of the 26 samples herein (see table S1 in supporting information (SI) for details). We classified these 6 power plants as Grade A (a typical plant is shown in figure 1(b)). Many other power plants are affected by nearby emitters, but leaving clear concentration gradients to perform a 2D Gaussian fitting analysis in sufficient unaffected directions (see figure 1(c)). We classified these 20 relatively isolated power plants as Grade B. We found that Fioletov et al's method [19] would lead to large uncertainties or even fail to converge when applied to Grade B power plants. Besides the nearby emitters, interference to a valid performance of the original 2D Gaussian fitting method also possibly arises from the complex topography around many target sources [32]. To estimate SO₂ burdens from these affected power plants under various complex background conditions, additional improvements must be made.

Zoom In Zoom Out Reset image size

We developed the SO₂ burden inversion method of Fioletov et al [19] in two aspects. First, benefiting from the removal of systematic offsets and latitude-dependent SO₂ column biases mentioned in section 2.2, a constant term, OMI_background, was added to the original 2D Gaussian function as a fitting parameter:

$$\begin{eqnarray}&&{{\rm{OMI}}}_{{{\rm{SO}}}_{{\rm{2}}}}={{\rm{OMI}}}_{{\rm{background}}}+\displaystyle \frac{a}{2\pi {\sigma }_{x}{\sigma }_{y}\sqrt{1-{\rho }^{2}}}\\ &&\quad \quad \quad \cdot \mathrm{exp}\left(-\displaystyle \frac{1}{2\left(1-{\rho }^{2}\right)}\right.\left[\displaystyle \frac{{\left(x-{\mu }_{x}\right)}^{2}}{{\sigma }_{x}^{2}}\right.\\ &&\quad \quad \quad +\left.\left.\displaystyle \frac{{\left(y-{\mu }_{y}\right)}^{2}}{{\sigma }_{y}^{2}}-\displaystyle \frac{2\rho \left(x-{\mu }_{x}\right)\left(y-{\mu }_{y}\right)}{{\sigma }_{x}{\sigma }_{y}}\right]\right),\end{eqnarray} \tag{ 1 }$$

where x and y indicate the coordinates of the OMI pixel center; the parameters μ_x and μ_y denote the position of the fitting maximum; the parameters σ_x, σ_y, and ρ determine the shape and rotation of the ellipse; and the parameter a represents the total number of SO₂ molecules (or SO₂ burden) observed by OMI associated with the target source. The local bias correction used in the original method, which subtracted the average SO₂ columns within a 300 km radius around the source, caused substantial signal degradation in the target domain (typically within 80 km) in this study, in view of the intensive emission hotspots in China. A similar constant background column term was also considered in McLinden et al [20]. By contract, the fitted OMI_background from the improved method in this study can better represent the background SO₂ columns compliant with the fitting domain surrounding each target source.

Second, a unique asymmetric fitting domain was customized for each selected coal-fired power plant according to the distribution of nearby sources of interfering emissions and of topography. Figure 1(c) shows a typical case for one Grade B power plant (#2 power plant). In this case, the northwest side of the target plant is less affected by other emission sources in contrast to the other directions, and a clear descent gradient of the SO₂ columns can be identified from the central source to the outmost grids in figure 1(c). Because the asymmetric fitting domains were used, instead of the uniform circular fitting domains (40 or 60 km radius depending on the source strength) in the original method, only OMI columns in directions that were less affected by surrounding interference were quantified in the following fitting; consequently, interference from nearby sources was minimized. A systematic bias may be introduced in the 2D Gaussian fitting for Grade B power plants due to the heterogeneous interference from the nearby emission sources at different parts of the fitting domain, which is difficult to quantify but can be partially canceled when analyzing the ratio of the fitted SO₂ burdens in two periods as conducted later in this study. Numerous OMI samples are required for the 2D Gaussian fitting so that sufficient fitting domain is needed for stable fitting results. Specially, we found that the reliability of the fitting was sensitively dependent on whether the background concentration areas were sufficiently covered by the asymmetric fitting domain. When this requirement was met, the biases in the 2D Gaussian fitting related to the domain choice, as mentioned by de Foy et al [23] can be greatly reduced and a stable solution can be thereby generated. Furthermore the biases in the 2D Gaussian fitting related to the dispersion directions, as mentioned by de Foy et al [23] have less effect on the analysis in this study, because no emissions were directly estimated herein and the systematic biases in the parameter a can be partially canceled in the assessment of trends. Typically, a domain with a width in excess of 60 km (from the central source to the farthest edge) was necessary in this study.

The SO₂ burdens over the selected 26 coal-fired power plants were estimated by applying the improved 2D Gaussian fitting method to the averaged OMI columns for the period 2006–2008, representing the strength of emissions before the effective operation of the FGD facilities as discussed in later section. The fitting parameter a for each target plant was then compared with the bottom-up SO₂ emissions, as shown in figure 2. These two sets of independent estimates are well correlated: R = 0.92 for all 26 samples, and R = 0.96 for the 6 Grade A samples, thereby indicating that the SO₂ emissions from isolated coal-fired power plants or similar large point sources can be linearly inferred using OMI measurements in China. Although the source strength of a single coal-fired power plant is not as large as that in the US and India, we found no statistically significant threshold on this factor in this study when using our improved fitting method and also the improved OMI SO₂ columns, in contrast with the findings of Fioletov et al [19] (>70 Gg yr⁻¹) and Lu et al [21] (>50 Gg yr⁻¹). The uncertainties of the inversed OMI SO₂ burdens were contributed by the biases in the OMI SO₂ slant and vertical column retrievals [27, 28, 33], the residual interference from other sources in the 2D Gaussian fitting method, and the relatively coarse resolution of the a priori SO₂ profiles used in the calculation of AMFs [25].

Zoom In Zoom Out Reset image size

The improved 2D Gaussian fitting method was further used to estimate the SO₂ burdens over these 26 coal-fired power plants before and after the effective operation of their FGD facilities. The results were then used to evaluate the real emission reductions during the period 2005–2012. It is widely believed that there is a long operation-vacuum period between the installation and effective operation dates for FGD facilities at coal-fired power plants in China due to the trade-off between the pressure of SO₂ emission reductions and the economic cost of desulfurization. The break points possibly began in 2008 to meet the nation's mandatory target of 10% SO₂ emission reductions in the '11th five-year plan' period (2006–2010), and they were partially due to the promotion of the Beijing 2008 Olympic Games. This circumstance has been portrayed by Li et al [16] for several coal-fired power plants in Inner Mongolia, China.

To identify the time of the FGD facilities begun operation for each of the 26 coal-fired power plants, we compared the temporal variations in the tropospheric SO₂ and NO₂ columns observed simultaneously by OMI. The NASA's Level-2 standard tropospheric NO₂ columns were used herein and processed following the same method as described in section 2.2 for SO₂ columns. The 0.01° × 0.01° SO₂ and NO₂ column data were then monthly averaged over a 0.1° × 0.1° region surrounding each power plant and 12-month moving smoothed to eliminate seasonal variations. Inter-annual variations in OMI NO₂ columns over isolated power plants can be used as an indicator for changes in activity rates of the sources. When FGD facilities came into effective operation, the emission factor of SO₂ largely decreased, so that the trend of OMI SO₂ columns can be significantly against that of the OMI NO₂ columns. This fact was examined in this work and used to determine the effective operation date (hereinafter referred to as the 'EOD') of the FGD facilities for each power plant herein. Similarly, we defined the scheduled operation date (usually at the time of installation) of the 'first FGD facility' at each power plant as the 'SOD'.

The 'SOD' was determined according to the scheduled operation date of the first FGD facility at the target power plant, which is provided by the bottom-up emission database. We then used the OMI NO₂ columns to constrain the trend of OMI SO₂ columns over these isolated power plants and identified when the OMI SO₂ columns varied significantly against the OMI NO₂ columns. The 'EOD' was determined by the ratio of the synchronous OMI SO₂ and NO₂ columns (12-month moving smoothed) over the target power plant. Figure 3(a) illustrates an example (for #3 power plant) for the determination of the EODs and SODs in this study. There were 4 generating units in this power plant, with the first FGD facility installed in November 2007. Therefore, the year 2007 was determined as the SOD for this power plant. The moving-averaged OMI SO₂/NO₂ has stably decreased significantly since the end of 2008 (figure 3(b)), reflecting the effective operation of the FGD facilities after that time. We defined the EOD point as the 12th month after the moving-average SO₂/NO₂ began to continuously decrease. We then identified the year of EOD point as the EOD year for this power plant (2009 for this case). For all 26 power plants in this study, the OMI-identified EODs and their SODs were listed in table S1. We found no EODs for these power plants are earlier than 2008 (see figure 4), and 20 power plants' EODs lagged behind their SODs. The average FGD lag period was determined to be 2.1 yr based on the average discrepancy between the EODs and SODs over the 26 power plants, which suggested that the real operating status of the FGD facilities was poor at the beginning of the '11th five-year plan' period (2006–2010).

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

We calculated the SO₂ burdens over these 26 coal-fired power plants for two time periods, before and after their EODs in 2005–2012 (the EOD year was included in the later period). The fitting parameters a for these two periods were compared with the corresponding bottom-up SO₂ emissions in figure 5, respectively. As shown in figure 5(a), the average SO₂ burdens for the 26 power plants were reduced by 38% after EODs, which is significantly less than 65% reduction estimated from the bottom-up inventory, indicating that the actual removal efficiencies of FGD facilities in these power plants may be lower than reported. We also found large difference in the correlations for these two periods, R = 0.91 before the EODs and R = 0.39 after the EODs (figure 5(b)). As stated in section 3.1, the good correlation before the EODs confirmed the accuracy of the OMI-derived SO₂ burden estimates, while the relatively poor correlation after the EODs further confirmed that power plants may have emitted more SO₂ than the expected levels estimated using the reported FGD removal efficiencies from the MEP.

Zoom In Zoom Out Reset image size

Since the construction of new units or the retirement of old units may also perturb the SO₂ emissions, the reduction rate of SO₂ burdens doesn't well represent the SO₂ removal efficiency of the FGD facilities. We then calculated the OMI-estimated SO₂ removal equivalence f_OMI for each power plant using the fitting parameters a for the periods before and after the individual EOD (a_before and a_after):

$$\begin{eqnarray}&&{f}_{{\rm{OMI}}}=\left(1-\displaystyle \frac{{a}_{{\rm{after}}}}{{a}_{{\rm{before}}}\cdot \left(1+k\right)}\right)\times 100\%,\end{eqnarray} \tag{ 2 }$$

where $k$ presents the increased ratio of activity level equivalence during the two periods, which is calculated as the increase of coal consumption rates. A negative value in k usually indicates the closedown of some generator units after the EOD at the corresponding power plant. By comparison, we calculated the estimated SO₂ removal equivalence using bottom-up SO₂ emission inventories with and without FGD operation at the target plants after the EODs. For all 26 power plants in this study, the SO₂ removal equivalences estimated from two approaches are listed in table S1. In weighted average, the SO₂ removal equivalence estimated from OMI was calculated to be 56.0 ± 21.3%, substantially lower than that from inventories (74.6 ± 9.1%) for the same period. To summarize all the findings we listed above, the real desulfurization efficiencies of FGD facilities at these 26 coal-fired power plants were considerably discounted because of the weak supervision measures in China.

It should be noted that other factors might also contribute to the differences between reported SO₂ emissions and SO₂ burdens. Uncertainties in SO₂ retrievals and fitting model could contribute to the errors. For example, the AMFs used in the satellite retrieving algorithm were impacted by the complex effects of aerosols [28], which depended on the relative vertical distribution of the aerosol layer and the tracer gas layer, the composition of the aerosols, and the wavelength used. The severe haze pollution in China may have complicated this issue. The systematic biases in the fitting model can be partially canceled when the ratio of the parameter a was used and would not lead to substantial uncertainties. Reduced SO₂ signals after EODs will increase the uncertainties in SO₂ retrieval and contribute to additional errors when comparing with emissions. The non-linear changes between changes in emissions and in SO₂ columns can also lead to uncertainties to our inverse model, although they were difficult to quantify.