Changes in Power Plant NO x Emissions over Northwest Greece Using a Data Assimilation Technique

: In this work, we investigate the ability of a data assimilation technique and space-borne observations to quantify and monitor changes in nitrogen oxides (NO x ) emissions over Northwestern Greece for the summers of 2018 and 2019. In this region, four lignite-burning power plants are located. The data assimilation technique, based on the Ensemble Kalman Filter method, is employed to combine space-borne atmospheric observations from the high spatial resolution Sentinel-5 Precursor (S5P) Tropospheric Monitoring Instrument (TROPOMI) and simulations using the LOTOS-EUROS Chemical Transport model. The Copernicus Atmosphere Monitoring Service-Regional European emissions (CAMS-REG, version 4.2) inventory based on the year 2015 is used as the a priori emissions in the simulations. Surface measurements of nitrogen dioxide (NO 2 ) from air quality stations operating in the region are compared with the model surface NO 2 output using either the a priori (base run) or the a posteriori (assimilated run) NO x emissions. Relative to the a priori emissions, the assimilation suggests a strong decrease in concentrations for the station located near the largest power plant, by 80% in 2019 and by 67% in 2018. Concerning the estimated annual a posteriori NO x emissions, it was found that, for the pixels hosting the two largest power plants, the assimilated run results in emissions decreased by ~40–50% for 2018 compared to 2015, whereas a larger decrease, of ~70% for both power plants, was found for 2019, after assimilating the space-born observations. For the same power plants, the European Pollutant Release and Transfer Register (E-PRTR) reports decreased emissions in 2018 and 2019 compared to 2015 ( − 35% and − 38% in 2018, − 62% and − 72% in 2019), in good agreement with the estimated emissions. We further compare the a posteriori emissions to the reported energy production of the power plants during the summer of 2018 and 2019. Mean decreases of about − 35% and − 63% in NO x emissions are estimated for the two larger power plants in summer of 2018 and 2019, respectively, which are supported by similar decreases in the reported energy production of the power plants (~ − 30% and − 70%, respectively).


Introduction
Emissions of nitrogen oxides (NO x = NO + NO 2 ) play a pivotal role in local and global atmospheric composition and air quality. NO x contributes to the formation of tropospheric ozone, peroxyacyl nitrate (PAN) and nitrate aerosols, and contributes to the environmental acidification [1]. Fossil fuel combustion, mainly originating from power plants, transport and industry, are the main anthropogenic sources of NO x in the atmosphere, while NO x is naturally emitted from soil, biomass burning and lightning [2]. Primarily, NO x is released

Power Plants in Northwest Greece
The Greek National Energy and Climate Plan (NCEP, https://ec.europa.eu/energy/ sites/ener/files/el_final_necp_main_en.pdf, accessed on 8 July 2021) is a well-developed strategy that provides environmental objectives, policies and measures. The NCEP integrates the targets set in the Directive 2016/2284/EC concerning the reduction of national emissions of certain atmospheric pollutants, such as NO x , SO 2 and NMVOC. Greece's commitment regarding NO x emissions is a reduction of 31% for the period between 2020 and 2029 and a reduction of 55% after 2030 compared to 2005. According to the European Pollutant Release and Transfer Register (E-PRTR), the region of western Macedonia in Northwest Greece, reported the largest NO x emissions from lignite power plants (about 60 kt) for 2007 followed by the Aegean Sea islands and Crete. NO x emissions in the region of Northwest Greece are reported to decrease significantly in the following years, declining to~17 kt in 2017. In this work, we hence focus on the regions of Northwest Greece, which hosts four of the largest lignite-burning power plants in Greece.
The topography of Northwest Greece is shown in Figure 1. In the centre of the area, a basin is located in 650 m above mean sea level, which is about~50 km long and 10 to 25 km wide, surrounded by mountains of around 1350 m above mean sea level. Inside the basin, small hills are present. Vegetation is restricted to isolated trees and small bushes [14]. Four lignite-burning power plants (annotated in Figure 1 21.59829 • E). In addition, one more lignite power plant, near the city of Bitola (~680 MW), is operating close to the border with the neighboring Republic of North Macedonia. The climate of the area is continental Mediterranean, characterized by high temperatures in summer and low temperatures in winter, while the prevailing winds during summer in the centre of the basin are mainly of NW, NNW and WNW directions (16.4, 15.6, and 9.9%, respectively) [14]. Nevertheless, the wind direction in the southern part of the basin is NE, dominated by the topography of the region. The villages and towns located in and around the basin, and their population are also shown in Figure 1. Finally, the locations of in situ air quality stations operating in the area are represented by black dots on the map.
The annual NO x 2015-2019 emissions, reported in E-PRTR based on measurements of the NO x mass concentrations, for the four large power plants in the region of Northwest Greece (dotted lines in Figure 2) are largest for the Ag. Dimitrios power plant, followed by Kardia, Amyntaio and Meliti, in accordance with their installed capacity. The energy production of the power plants (solid lines in Figure 2), reported by the Energy Exchange Group-EnEx (www.enexgroup.gr, accessed on 14 April 2021), follows the variability of the emissions throughout the years, showing large decreases in energy production in 2019 mainly in the Ag. Dimitrios and Kardia plants. The important reduction of NO x emissions in the region is further confirmed by NO 2 column measurements from OMI aboard of the EOS-Aura satellite [15]. The deseasonalized monthly tropospheric NO 2 columns at the satellite pixel where the large power plant of Ag. Dimitrios is located is shown in Figure S1 (top) between 2005 and 2020. Especially after 2016, strong reduction in NO 2 levels is seen and corresponds very well with the reported decreases in Figure 2. The mean annual tropospheric NO 2 columns are shown in Figure S1 (bottom) together with the corresponding trend showing a decrease of about 2.5 × 10 15 molec cm −2 (about −55%) per decade. Since the emissions reported by independent sources show an important decrease, even between the years 2018 and 2019, we study here the ability of the assimilation of satellite observations to sense these changes as well. in the region is further confirmed by NO2 column measurements from OMI aboard of the EOS-Aura satellite [15]. The deseasonalized monthly tropospheric NO2 columns at the satellite pixel where the large power plant of Ag. Dimitrios is located is shown in Figure  S1 (top) between 2005 and 2020. Especially after 2016, strong reduction in NO2 levels is seen and corresponds very well with the reported decreases in Figure 2. Τhe mean annual tropospheric NO2 columns are shown in Figure S1 (bottom) together with the corresponding trend showing a decrease of about 2.5 × 10 15 molec cm −2 (about −55%) per decade. Since the emissions reported by independent sources show an important decrease, even between the years 2018 and 2019, we study here the ability of the assimilation of satellite observations to sense these changes as well.

The LOTOS-EUROS CTM
The LOTOS-EUROS (Long Term Ozone Simulation-EURopean Operational Smog model) (https://lotos-euros.tno.nl/, accessed on 7 July 2021) simulates the air pollution in the troposphere while a detailed description of the model is available in [16]. LOTOS-EUROS is one of the nine state-of-the-art systems used in the operational Copernicus Atmosphere Monitoring Services (CAMS, https://atmosphere.copernicus.eu/, accessed on 7 July 2021).
In this study, we use the model version 2.2.001. The simulations cover the domain in Northwest Greece ( Figure 3) and expand from 40.2° to 41.2°N and 21.2° to 22.2°E. The horizontal spatial resolution is set to 0.1° longitude × 0.05° latitude (about 10 km × 5 km at Greece latitudes). In the vertical, ten hybrid sigma-pressure layers are used with a top at about 200 hPa; these are obtained as a coarsening of the layers in the meteorological input. The gas phase chemistry follows a modified version of Carbon Bond Mechanism IV scheme (CBM-IV) [17] while the aerosol chemistry uses the ISORROPIA II parameterization [18]. Meteorological variables are obtained at 7 km × 7 km spatial resolution from the Integrated Forecasting System (IFS) of the European Centre for Medium-Range Weather Forecasts (ECMWF). The surface-layer meteorological variables are obtained at hourly resolution while the 3-dimensional variables at model levels every 3 h [19]. Boundary and initial conditions are obtained from the Copernicus Atmosphere Monitoring Service

The LOTOS-EUROS CTM
The LOTOS-EUROS (Long Term Ozone Simulation-EURopean Operational Smog model) (https://lotos-euros.tno.nl/, accessed on 7 July 2021) simulates the air pollution in the troposphere while a detailed description of the model is available in [16]. LOTOS-EUROS is one of the nine state-of-the-art systems used in the operational Copernicus Atmosphere Monitoring Services (CAMS, https://atmosphere.copernicus.eu/, accessed on 7 July 2021).
In this study, we use the model version 2.2.001. The simulations cover the domain in Northwest Greece ( Figure 3) and expand from 40.2 • to 41.2 • N and 21.2 • to 22.2 • E. The horizontal spatial resolution is set to 0.1 • longitude × 0.05 • latitude (about 10 km × 5 km at Greece latitudes). In the vertical, ten hybrid sigma-pressure layers are used with a top at about 200 hPa; these are obtained as a coarsening of the layers in the meteorological input. The gas phase chemistry follows a modified version of Carbon Bond Mechanism IV scheme (CBM-IV) [17] while the aerosol chemistry uses the ISORROPIA II parameterization [18]. Meteorological variables are obtained at 7 km × 7 km spatial resolution from the Integrated Forecasting System (IFS) of the European Centre for Medium-Range Weather Forecasts (ECMWF). The surface-layer meteorological variables are obtained at hourly resolution while the 3-dimensional variables at model levels every 3 h [19]. Boundary and initial conditions are obtained from the Copernicus Atmosphere Monitoring Service (CAMS, https://atmosphere. copernicus.eu/, accessed on 7 July 2021), global near-real time (NRT) product, at a 3-hour temporal resolution and a spatial resolution of about 35 km. Anthropogenic emissions are taken from the CAMS-REG (CAMS-Regional European emissions) inventory version 4.2 for the year 2015 [20] at a spatial resolution of 0.1 • longitude and 0.05 • latitude while the emissions temporal profiles used are the default provided with the inventory. The latest available year in this inventory is 2015, and this year is therefore used for the simulations in this study. The emissions are vertically distributed in the model based on fixed height profiles following the approach of EURODELTA project [21]. The emissions are distributed over 8 vertical layers depending on the source category and type (area or point source). In particular, power plants emit from high stacks and more than 90% is considered to be emitted between 184 m and 552 m. The distribution from the height layers to the model layers is hence computed. Biogenic emissions are calculated online using actual meteorology and depend on a detailed land use and tree-species database described in [22]. Soil NO emissions are taken from a parametrization depending on soil type and soil temperature [23]. Emissions for lightning are not included in the simulations. Emissions from biomass burning are obtained from the Global Fire Assimilation System (GFAS) dataset [24].  [20] at a spatial resolution of 0.1 • longitude and 0.05 • latitude while the emissions temporal profiles used are the default provided with the inventory. The latest available year in this inventory is 2015, and this year is therefore used for the simulations in this study. The emissions are vertically distributed in the model based on fixed height profiles following the approach of EURODELTA project [21]. The emissions are distributed over 8 vertical layers depending on the source category and type (area or point source). In particular, power plants emit from high stacks and more than 90% is considered to be emitted between 184 m and 552 m. The distribution from the height layers to the model layers is hence computed. Biogenic emissions are calculated online using actual meteorology and depend on a detailed land use and tree-species database described in [22]. Soil NO emissions are taken from a parametrization depending on soil type and soil temperature [23]. Emissions for lightning are not included in the simulations. Emissions from biomass burning are obtained from the Global Fire Assimilation System (GFAS) dataset [24].  An extensive evaluation of the overall LOTOS-EUROS performance over Greece using ground-based measurements and satellite derived observations has been performed by [25]. In summary, it has been found that the modelled NO2 columns show a high spatial correlation (0.95) and a negative bias of −18% when compared with S5P/TROPOMI tropospheric columns over Athens in the summertime. The a priori emission inventory and boundary layer height assumptions were found to be significant sources of uncertainty in  An extensive evaluation of the overall LOTOS-EUROS performance over Greece using ground-based measurements and satellite derived observations has been performed by [25]. In summary, it has been found that the modelled NO 2 columns show a high spatial correlation (0.95) and a negative bias of −18% when compared with S5P/TROPOMI tropospheric columns over Athens in the summertime. The a priori emission inventory and boundary layer height assumptions were found to be significant sources of uncertainty in model simulations.

The A Priori NOx Emissions
The total a priori anthropogenic and biogenic emissions, used for the simulations in the summer of 2019, are shown in Figure 3. The power plant locations are identified by the green star symbols. The black triangles represent cities, the size of the triangle is proportional to the population.
The sum of the anthropogenic emissions in the grid-cells where the power plants are located constitute more than 95% of the total emissions (including biogenic), making anthropogenic the primary NO x source in this region ( Figure 3). The annual anthropogenic emissions from the point sources of public power sector in CAMS-REG over the pixels where the three larger power plants are located are by far dominant (>96%) compared to the rest emitting anthropogenic sources (Table 1). In the grid pixel where the smaller capacity, Meliti, power plant is located, the rest of the emitting anthropogenic emissions account for about 15% of the total anthropogenic emissions.

The S5P/TROPOMI Satellite Observations
The Sentinel-5 Precursor (S5P) satellite, launched in October 2017, carries the TRO-POsperic Monitoring Instrument (TROPOMI), a passive nadir-viewing spectrometer. TRO-POMI provides measurements of the atmospheric composition at an unprecedented spatial resolution of 7 × 3.5 km 2 at nadir (5.5 × 3.5 km 2 since 6 August 2019) with near global coverage in one day. S5P is in an orbit at an altitude of 817 km with an overpass of around 13:30 local solar time [26]. The TROPOMI NO 2 retrieval algorithm is based on the DOMINO NO 2 retrieval that is used on its predecessor instrument OMI/Aura [27] and is developed by the Royal Netherlands Meteorological Institute (KNMI), described in the product Algorithm Theoretical Basis Document (ATBD, [28]). Validation studies indicated that TROPOMI systematically underestimates the NO 2 columns at extremely polluted regions reaching up to biases of 30% [29][30][31], while it overestimates low NO 2 columns [32,33]. Recent studies [29,32] suggest that the use of high-resolution a priori profiles from regional models (e.g., from the regional CAMS ensemble) will increase the retrieved tropospheric column by 10-30% over polluted locations with high emissions, explaining part of the bias. Notably, when model outputs are compared with TROPOMI observations using the averaging kernels (as is done in this paper), the a priori profile shape used in the retrieval does not influence the relative comparison. Other uncertainties are related to the surface albedo climatology, and cloud (aerosol) retrievals, and may be responsible for remaining biases of order 10-20%. The TROPOMI NO 2 data are routinely validated by comparisons to ground-based reference measurements by the Mission Performance Center Validation Data Analysis Facility (VDAF, http://mpc-vdaf.tropomi.eu, accessed on 7 July 2021).
The NO 2 S5P/TROPOMI observations used for the assimilation in the present study are obtained via the Copernicus Open Data Access Hub (https://s5phub.copernicus.eu/, accessed on 7 July 2021). The level 2 reprocessed data, RPRO, v01.02 for the summer of 2018, while for 2019 the offline data (OFFL v1.03) are used. The data are filtered for a quality assurance value higher than 0.75 to ensure mostly cloud-free pixels, as recommended by the Product User Manual (PUM, [34]).

Ensemble Kalman Filter around LOTOS-EUROS CTM
As mentioned above, the emissions used in the simulation model form a source of uncertainty. In the context of the data assimilation algorithm, the uncertain emissions are modelled according to: where e b [k] are the a priori emissions at time k and δe[k] is a stochastic emission correction factor. In order to specify a smooth uncertainty as described in [35], the emission correction factor follows a structure of a colored noise process instead of a white noise that would be uncorrelated in time [36]. The colored noise has zero mean and standard deviation σ and is implemented by: where w[k] is a white noise vector uncorrelated in time with zero mean and unity standard deviation. Factor α ∈ [0,1] is a time correlation parameter that is used to describe the temporal variation following: where τ is a temporal length scale, and t[k] − t[k − 1] is the time step used. The choice of a suitable value for τ is described in Appendix A.
In the Ensemble Kalman Filter, the model uncertainty is represented by an ensemble of state vectors. Here, an augmented state vector is used, containing both the concentrations (c) and the emissions correction factors ( δe). In this way, the assimilated states also provide an estimate of the emission correction. The state vectors are propagated in time using the LOTOS-EUROS model, M, and the colored noise model: The ensemble is propagated by the model until observations, y[k], become available. The observations are then used to analyze the ensemble such that the remain distribution is in agreement with the observations. The simulation of the observation by a state x is described by: where H is the nonlinear observation operator used to transform the background states from the model space to the observational space; for the satellite NO 2 observations this contains the averaging kernel (AK) of the satellite product. Vector v[k] is the observation representation error with zero mean and covariance R. The covariance R is defined by the retrieval errors provided by the observational data. For the analysis the Local Ensemble Transform Kalman Filter-LETKF algorithm is used [37]. The algorithm analyses the state per grid cell. First, all the observations and the corresponding ensemble simulations within a certain distance from the grid cell are collected. This distance depends on the selection of a length-scale called localization radius, ρ. Observations within a distance of 3.5ρ from the grid cell are selected and weighted relative to the inverse of their distance from the grid cell and based on the ρ. These are then used to update the ensemble in the grid cell. For small length-scales, the analysis only changes the ensemble in the grid cells with or close to observations, while using a longer length-scale more observations are used for the analysis of a single grid cell. In this way, the observations that affect the point analysis in each time step depend on the length-scale. In this study, we defined a localization radius equal to 14 km following sensitivity experiments described in Appendix A. The required ensemble size depends (among others) on the choices for ρ and τ, the sensitivity experiments suggested that a small ensemble size of only 12 members is sufficient for the chosen configuration.

In Situ NO 2 Measurements
For the validation of the simulations, as well as for choosing the optimal configuration of the assimilation system over the studied region, hourly in situ surface NO 2 measurements from air quality stations in the region have been acquired for the summers of 2018 and 2019 (Dr Evagelopoulos V., University of Western Macedonia, private communication). The location of the stations is shown in Figure 1

LOTOS-EUROS NO2 Simulations and S5P/TROPOMI Observations
The S5P/TROPOMI tropospheric NO2 observations are regridded onto the model grid (0.10° × 0.05°) using a standard area-weighted averaging method. The observations are weighted depending on their distance from each grid point and the relative average is then calculated. Simulations of the regridded observations are then obtained by application of the AK of the satellite product to the LOTOS-EUROS concentrations. The simulated concentrations are first mapped to the retrieval's a priori layers to make the comparison between them feasible; the model top at 200 hPa is exactly sufficient to cover the TROPOMI tropospheric column. In this way, the vertical sensitivity of the satellite instrument is taken into account in the comparisons between the model and the observations. The averaging kernels are applied to the model output at the closest time of the satellite overpass, which is on average between 11:00 and 12:00 UTC.
For the summer of 2019, the averaged NO2 tropospheric columns as observed by S5P/TROPOMI are shown in Figure 5a. The corresponding model NO2 values after the AK of the satellite retrieval are applied are shown in Figure 5b. The same figures for the summer of 2018 are shown in Figure S4. The number of available daily satellite observations over the two larger power plants is lower in 2018 compared to 2019, with ~53 and ~70 pixels, respectively. The retrievals and the simulations show similar spatial distributions, with higher values around the two largest power plants, Ag. Dimitrios and Kardia. However, large discrepancies in the absolute values of NO2 are found around the larger power plants where LOTOS-EUROS overestimates NO2 in comparison with S5P/TRO-POMI observations. During the summer of 2019, LOTOS-EUROS simulates an average column of about 9 × 10 15 molec cm −2 while TROPOMI observes about 2 × 10 15 molec cm −2 over the same grid cell where Ag. Dimitrios is located ( Table 2). A high model bias is also found over the power plant of Kardia (bias of 4.5 × 10 15 molec cm −2 ), whereas the discrepancy is negligible over the small Meliti power plant. The column simulations are lower for the summer of 2018 than for the summer of 2019 ( Figure 6) even though the a priori

LOTOS-EUROS NO 2 Simulations and S5P/TROPOMI Observations
The S5P/TROPOMI tropospheric NO 2 observations are regridded onto the model grid (0.10 • × 0.05 • ) using a standard area-weighted averaging method. The observations are weighted depending on their distance from each grid point and the relative average is then calculated. Simulations of the regridded observations are then obtained by application of the AK of the satellite product to the LOTOS-EUROS concentrations. The simulated concentrations are first mapped to the retrieval's a priori layers to make the comparison between them feasible; the model top at 200 hPa is exactly sufficient to cover the TROPOMI tropospheric column. In this way, the vertical sensitivity of the satellite instrument is taken into account in the comparisons between the model and the observations. The averaging kernels are applied to the model output at the closest time of the satellite overpass, which is on average between 11:00 and 12:00 UTC.
For the summer of 2019, the averaged NO 2 tropospheric columns as observed by S5P/TROPOMI are shown in Figure 5a. The corresponding model NO 2 values after the AK of the satellite retrieval are applied are shown in Figure 5b. The same figures for the summer of 2018 are shown in Figure S4. The number of available daily satellite observations over the two larger power plants is lower in 2018 compared to 2019, with~53 and~70 pixels, respectively. The retrievals and the simulations show similar spatial distributions , with higher values around the two largest power plants, Ag. Dimitrios and Kardia. However, large discrepancies in the absolute values of NO 2 are found around the larger power plants where LOTOS-EUROS overestimates NO 2 in comparison with S5P/TROPOMI observations. During the summer of 2019, LOTOS-EUROS simulates an average column of about 9 × 10 15 molec cm −2 while TROPOMI observes about 2 × 10 15 molec cm −2 over the same grid cell where Ag. Dimitrios is located (Table 2). A high model bias is also found over the power plant of Kardia (bias of 4.5 × 10 15 molec cm −2 ), whereas the discrepancy is negligible over the small Meliti power plant. The column simulations are lower for the summer of 2018 than for the summer of 2019 ( Figure 6) even though the a priori emissions used are the same. This can be attributed to the different meteorological conditions taking place that affect the NO 2 levels together with the different number of daily available satellite observations.

Updated Assimilated NOx Emissions
The choice for localization radius ρ and the temporal parameter τ is crucial for the performance of the assimilation system, and are studied with different sensitivity tests

Updated Assimilated NOx Emissions
The choice for localization radius ρ and the temporal parameter τ is crucial for the performance of the assimilation system, and are studied with different sensitivity tests

Updated Assimilated NO x Emissions
The choice for localization radius ρ and the temporal parameter τ is crucial for the performance of the assimilation system, and are studied with different sensitivity tests shown in Appendix A. The selected localization radius is 14 km and the temporal parameter is 7 days, since these values showed the best performance during the assimilation. A rather small ensemble size of 12 members was sufficient to obtain stable results, as also found in other studies using the same system [38].
After assimilation, the LOTOS-EUROS columns are decreased as expected ( Figure 6). The biases between the assimilated columns and the observations over the four power plants are lower than 0.6 molec cm −2 for both 2018 and 2019. In the grid cell where Ag. Dimitrios is located, the average NO 2 column in the summer of 2019 (2018) in the assimilated run is about 2.4 × 10 15 molec cm −2 (4 × 10 15 molec cm −2 ), while TROPOMI observes about 2 × 10 15 molec cm −2 (3.70 × 10 15 molec cm −2 ) during the same period.
The a priori and a posteriori NO x emissions during the summer of 2019 over the four grid pixels where the power plants are located are shown in Figure S5 for 2018 (top) and 2019 (bottom). The TROPOMI-based emissions are in general lower relative to the a priori emissions which refer to year 2015. The updated emissions over the power plant of Ag. Dimitrios are 38% and 63% lower than a priori emissions in 2018 and 2019, respectively (Table 3). Similarly, over the Kardia power plant the emissions were decreased by 27% in 2018 and by 63% in 2019. In 2018, over the Amyntaio plant negligible emission changes were inferred whereas in 2019 the a posteriori emissions are 37% lower than the a priori. Finally, over the smallest plant of Meliti no important differences are found, and only a small decrease of 11% is found for 2018 (Table 3). Table 3. Relative differences between a posteriori emissions estimated in the summer of 2018 and 2019 and a priori emissions used for the simulations over the four grid-pixels.

Relative Differences
Ag. Dimitrios Kardia Amyntaio Meliti The daily time series of NO x emissions over the grid cell where Ag. Dimitrios is located is shown in top of Figure 7 for 2018 (left) and 2019 (right). A clear weekly profile is visible in the a priori emissions (orange line) representing the lower NO x emissions assumed during the weekends and stable emissions during the week. It is also shown that during July the emissions are slightly lower than in June and August (~−1% and~−8%, respectively). However, the a posteriori emissions (purple line) do not follow these profiles. In 2018 the estimated emissions are found to strongly decrease in August by around 40% and 30% compared to June and July, respectively while in 2019 they remain low throughout the period of summer and decrease by 28% in August compared to June and July. A clear day-of-the week profile is not visible either. A large τ selection is reasonable for the case of the power plants, since a large day-to-day emission variation is not expected, but rather a relatively constant bias during all months. No changes in the diurnal variability are found (bottom of Figure 7), as could be expected, since only one satellite measurement per day is available. However, the a posteriori diurnal emissions in 2019 are much lower than the inferred emissions in 2018 and are decreased by more than 60% compared to the a priori emissions. The diurnal variability of NO x emissions of the rest of the power plants is given in Figure S6. well. The case of Meliti power plant is less representative of the grid cell where it is located, since as already discussed before (in Section 2.2) there are more emission sources in the region and the power plant has a small capacity (~290 MW). Moreover, in Figure 5 a first indication of transboundary pollution was evident. To further prove this, wind roses are plotted using the ECMWF wind components used for the simulations, for the hour closer to the TROPOMI overpass over the area (~12:00 UTC). Figure S7 and Figure S8 show the wind direction and speed over the pixels of the power plant and the city of Bitola, in the neighboring country (Republic of North Macedonia), for the summer of 2018 and 2019, respectively. The dominant winds for both seasons are North and Northeast. Furthermore, Figure S9 shows that the dominant winds over Meliti power plant are of a Northwest direction for both 2018 and 2019. According to these findings, it is safe to assume that NO x from Bitola is affecting the Meliti power plant near the border of the two countries.

In Situ Measurements
In order to validate the a posteriori NOx simulations, surface NO2 concentrations simulated using the a priori emissions (base run) and the TROPOMI-based emissions (assim-

In Situ Measurements
In order to validate the a posteriori NO x simulations, surface NO 2 concentrations simulated using the a priori emissions (base run) and the TROPOMI-based emissions (assimilated run) are compared with hourly in situ NO 2 measurements. In Figure 9, the mean summer surface NO 2 as simulated by the base run (left) and the assimilated run (right) are shown together with the mean summer NO 2 values of the in situ stations in 2019 ( Figure S10 for 2018), depicted as colored circles. The assimilated NO 2 concentrations are reduced over the whole region and especially around the power plants of Ag. Dimitrios and Kardia compared to the base run (Figure 9 and Figure S10).

In Situ Measurements
In order to validate the a posteriori NOx simulations, surface NO2 concentrations simulated using the a priori emissions (base run) and the TROPOMI-based emissions (assimilated run) are compared with hourly in situ NO2 measurements. In Figure 9, the mean summer surface NO2 as simulated by the base run (left) and the assimilated run (right) are shown together with the mean summer NO2 values of the in situ stations in 2019 ( Figure  S10 for 2018), depicted as colored circles. The assimilated NO2 concentrations are reduced over the whole region and especially around the power plants of Ag. Dimitrios and Kardia compared to the base run (Figures 9 and S10).   (Table S1 and Table 4). Furthermore, the diurnal variability of the air quality observations in Koilada ( Figure 10 and Figure S11 left) is better represented by the assimilated runs since the concentrations are much closer to the measurements, as is also found for the morning diurnal simulations at the Florina air quality station ( Figure 10 and Figure S11 right).
Similarly, surface NO 2 concentrations measured from the Amyntaio air quality station, which is affected by the neighboring town and power plant, are lower than the base run results (bias 3.49 µg/m 3 in 2019 and 1.13 µg/m 3 in 2018), while the biases are reduced when compared with the assimilated concentrations (bias 1.73 µg/m 3 and 0.85 µg/m 3 in 2019 and 2018, respectively). It is also worth mentioning that the NO 2 measurements in 2019 are lower than in 2018 in this case as well (4.00 ± 2.25 µg/m 3 and 6.37±3.82 µg/m 3 , respectively).
Surface NO 2 levels measured in Meliti station are high in 2019 (8.18 ± 6.46 µg/m 3 ) and 2018 (10.93 ± 11.05 µg/m 3 ) and the model underestimates the NO 2 levels both before and after assimilation. This is possibly due to possible transboundary pollution from the Bitola power plant, already discussed in the literature [39]. To state that, as already discussed in Section 2.5, the prevailing winds of N-NW direction, appear to transport pollution from the power plant and the city of Bitola in Republic of North Macedonia to North Greece. to the measurements, as is also found for the morning diurnal simulations at the Florina air quality station (Figures 10 and S11 right). Similarly, surface NO2 concentrations measured from the Amyntaio air quality station, which is affected by the neighboring town and power plant, are lower than the base run results (bias 3.49 μg/m 3 in 2019 and 1.13 μg/m 3 in 2018), while the biases are reduced when compared with the assimilated concentrations (bias 1.73 μg/m 3 and 0.85 μg/m 3 in 2019 and 2018, respectively). It is also worth mentioning that the NO2 measurements in 2019 are lower than in 2018 in this case as well (4.00 ± 2.25 μg/m 3 and 6.37±3.82 μg/m 3 , respectively). Surface NO2 levels measured in Meliti station are high in 2019 (8.18 ± 6.46 μg/m 3 ) and 2018 (10.93 ± 11.05 μg/m 3 ) and the model underestimates the NO2 levels both before and after assimilation. This is possibly due to possible transboundary pollution from the Bitola power plant, already discussed in the literature [39]. To state that, as already discussed in Section 2.5, the prevailing winds of N-NW direction, appear to transport pollution from the power plant and the city of Bitola in Republic of North Macedonia to North Greece.

Energy Production of the Power Plants
In order to further evaluate the updated a posteriori NOx emissions, the reported monthly energy production data per power station are examined. The aforementioned  (Figure 11b). It is also shown that energy production and NO x emissions in 2019 are even lower than in 2018. The relative differences between the a priori and a posteriori emissions directly compared to the relative differences between the power plant energy productions are shown in Table 5  and 2019. Larger decreases are found for both energy and emissions in 2019 compared to 2018 as well. In the case of the Meliti power plant, the emissions remain nearly the same, despite the fact that the energy production in 2019 is twice lower than in 2015 and 2018. Overall, the emission reduction is even more pronounced in 2019 compared to 2015 at the three largest power plants of Ag. Dimitrios, Kardia and Amyntaio and this is further confirmed by the reduction of the energy production in the same year that reaches at least 50%.

Conclusions
In this work, an advanced data assimilation system is used to estimate the NOx emission changes at four lignite power plants operating in Northwest Greece. An ensemble Kalman filter operating around the LOTOS-EUROS CTM is employed to estimate the NOx surface emissions with a resolution of 0.1° in longitude and 0.05° in latitude using

Conclusions
In this work, an advanced data assimilation system is used to estimate the NO x emission changes at four lignite power plants operating in Northwest Greece. An ensemble Kalman filter operating around the LOTOS-EUROS CTM is employed to estimate the NO x surface emissions with a resolution of 0.1 • in longitude and 0.05 • in latitude using S5P/TROPOMI NO 2 column retrievals. The a priori anthropogenic emissions are based on reports for 2015 and are obtained from Copernicus Atmosphere Monitoring Service-Regional European emissions (CAMS-REG) version 4.2. Our main conclusions can be summarized below. Overall, the method proposed here is appropriate in detecting emission trends of local large emitters, and could be valuable also for regions in the world where no up-to-date emission inventories are available. The results of this study support the idea that NO 2 observations of high temporal and spatial resolution can be used in data assimilation of a validated CTM via the use of the EnKF in order to improve NO x emissions. Funding: This research has been co-financed by the European Union (European Regional Development Fund) and Greek national funds through the Operational Program "Competitiveness, Entrepreneurship and Innovation" (NSRF 2014-2020) by the "Panhellenic Infrastructure for Atmospheric Composition and Climate Change" project (MIS 5021516) and well as the "Innovative system for Air Quality Monitoring and Forecasting" project [code T1EDK-01697, MIS 5031298), implemented under the Action "Reinforcement of the Research and Innovation" Infrastructure. Since different parameters may affect the performance of the assimilation system, a series of sensitivity tests were conducted in order to select the optimal configuration for the present study. The experiments were performed for July 2019 and the NO 2 surface simulations from the base run, using the a priori NO x emissions, and the distinct assimilation experimental runs, using the updated a posteriori NO x emissions, are compared with hourly NO 2 in situ measurements of the air quality station Koilada (Table A1). The in situ station Koilada is located at the same grid cell as the Ag. Dimitrios power plant, which is the pixel with the largest bias between the model predictions and the S5P/TROPOMI observations (Table A1). Table A1. Performance of the EnKF system for different parameters; the localization radius (ρ) and the temporal correlation parameter (τ). Hourly in situ measurements in the Koilada air quality station are compared with hourly surface simulations for July month. The statistics are given in µg/m 3 . The first parameter examined is the localization radius (ρ) described in Section 2.4. The observations that affect the grid cell analysis in each time step depend on the length scale, which means that the larger the length selected the more observations are used for the analysis of a single grid cell. According to [40] a localization procedure will prevent spurious correlations that may appear in the analysis due to the finite ensemble size selected. For the sensitivity tests, a temporal correlation parameter was set equal to 3 days and 5 different values of ρ were examined: ρ of 0, 5, 7, 14 and 20 km. The performance improves when the length is increased to 14 km and the bias decreases to 2.68 µg/m 3 (Table A1), compared to a bias of 12.21 µg/m 3 of the base run and 6.01 µg/m 3 when selecting ρ equal to 7 km. The bias does not improve significantly when the localization radius is further increased to 20 km (bias of 2.61 µg/m 3 ) ( Figure A1). Moreover, since this study is focused on point sources of power plants and in the area around the stations no other significant NO x sources are found, choosing a very large length may affect the results with observations of background values leading to an artificial NO x decrease. The first parameter examined is the localization radius (ρ) described in Section 2.4. The observations that affect the grid cell analysis in each time step depend on the length scale, which means that the larger the length selected the more observations are used for the analysis of a single grid cell. According to [40] a localization procedure will prevent spurious correlations that may appear in the analysis due to the finite ensemble size selected. For the sensitivity tests, a temporal correlation parameter was set equal to 3 days and 5 different values of ρ were examined: ρ of 0, 5, 7, 14 and 20 km. The performance improves when the length is increased to 14 km and the bias decreases to 2.68 μg/m 3 (Table A1), compared to a bias of 12.21 μg/m 3 of the base run and 6.01 μg/m 3 when selecting ρ equal to 7 km. The bias does not improve significantly when the localization radius is further increased to 20 km (bias of 2.61 μg/m 3 ) ( Figure A1). Moreover, since this study is focused on point sources of power plants and in the area around the stations no other significant NOx sources are found, choosing a very large length may affect the results with observations of background values leading to an artificial NOx decrease. The second sensitivity experiment performed concerns the temporal correlation parameter (τ) described in Section 2.4. Different values of τ are studied in the assimilation system; τ of 1, 3, 5 and 7 days, while the localization radius during the tests is fixed and equal to 14 km. The overestimations of the NO 2 simulations when compared to the in situ measurements reduces to a bias of 1.57 µg/m 3 (Table A1) when τ is equal to 7 days while when τ is small the NO 2 concentrations are less influenced and closer to the base run simulations ( Figure A2). As a result, the temporal parameter used for the assimilations is set to 7 days.

Localization Radius (ρ) in km
here 2021, 12, x FOR PEER REVIEW 19 of 21 system; τ of 1, 3, 5 and 7 days, while the localization radius during the tests is fixed and equal to 14 km. The overestimations of the NO2 simulations when compared to the in situ measurements reduces to a bias of 1.57 μg/m 3 (Table A1) when τ is equal to 7 days while when τ is small the NO2 concentrations are less influenced and closer to the base run simulations ( Figure A2). As a result, the temporal parameter used for the assimilations is set to 7 days.