Time-resolved emission reductions for atmospheric chemistry modelling in Europe during the COVID-19 lockdowns

We quantify the reductions in primary emissions due to the COVID-19 lockdowns in Europe. Our estimates are provided in the form of a dataset of reduction factors varying per country and day that will allow modelling and identifying the associated impacts upon air quality. The countryand daily-resolved reduction factors are provided for each of the following source categories: energy industry (power plants), manufacturing industry, road traffic and aviation (landing and take-off cycle). We computed the reduction factors based on open access and near-real time measured activity data from a 15 wide range of information sources. We also trained a machine learning model with meteorological data to derive weathernormalised electricity consumption reductions. The time period covered is from 21 February, when the first European localised lockdown was implemented in the region of Lombardy (Italy), until 26 April 2020. This period includes five weeks (23 March until 26 April) with the most severe and relatively unchanged restrictions upon mobility and socio-economic activities across Europe. The computed reduction factors were combined with the Copernicus Atmosphere Monitoring Service’s European 20 emission inventory using adjusted emission temporal profiles in order to derive time-resolved emission reductions per country and pollutant sector. During the most severe lockdown period, we estimate the average emission reductions to be -33% for NOx, -8% for NMVOC, -7% for SOx and -7% for PM2.5 at the EU-30 level (EU-28 plus Norway and Switzerland). For all pollutants more than 85% of the total reduction is attributable to road transport, except SOx. The reductions reached -50% (NOx), -14% (NMVOC), -12% (SOx) and -15% (PM2.5) in countries where the lockdown restrictions were more severe such 25 as Italy, France or Spain. To show the potential for air quality modelling we simulated and evaluated NO2 concentration decreases in rural and urban background regions across Europe (Italy, Spain, France, Germany, United-Kingdom and Sweden). We found the lockdown measures to be responsible for NO2 reductions of up to -58% at urban background locations (Madrid, Spain) and -44% at rural background areas (France), with an average contribution of the traffic sector to total reductions of 86% and 93%, respectively. A clear improvement of the modelled results was found when considering the emission reduction 30 factors, especially in Madrid, Paris and London where the bias is reduced with more than 90%. Future updates will include the extension of the COVID-19 lockdown period covered, the addition of other pollutant sectors potentially affected by the https://doi.org/10.5194/acp-2020-686 Preprint. Discussion started: 22 July 2020 c © Author(s) 2020. CC BY 4.0 License.

Considering all of the above, the quantification of emission changes due to the COVID-19 lockdown requires the use of 65 reduction factors that are, at least: (i) country-dependent, (ii) pollutant sector-dependent and (iii) daily dependent for some sectors. Some studies focussing on the quantification of emission reductions are beginning to be published. Le Quéré et al. (2020) quantified the reduction in daily CO2 emissions during the COVID-19 lockdown from January 2020 to April 2020 over 69 countries, 50 US states and 30 Chinese provinces for a total of six sectors of the economy (i.e. energy industry, manufacturing industry, road transport, residential sector, public sector and aviation). The study, which calculates the emission 70 reductions based on national activity data, was focussed on estimating the expected impact of the lockdowns upon the 2020 annual CO2 emissions and climate, but it did not include an analysis of emission cuts of criteria pollutants (NOx, SOx, NMVOC, NH3, PM10 and PM2.5) or air pollution levels. More recently, Menut et al. (2020) developed an emission scenario for Western Europe to quantify the impact of the lockdowns on air quality levels. Although focussing on criteria pollutants, the emission scenario was limited to March 2020 and was set up using only the Apple movement trends, which were used to derived 75 emission reductions not only for road transport but also for other anthropogenic sources (i.e. manufacturing industry, non-road transport and residential/commercial combustion).
We present an open-source dataset of daily, sector-and country-dependent emission reduction factors for Europe associated with the COVID-19 lockdowns. These factors are designed to both support the quantification of European primary emission 80 reductions and the associated impacts upon air quality. Our emission reduction factors are based on a bottom-up approach that considers a wide range of information sources, including open access and near-real time measured activity data, proxy indicators and other available reports. The resulting dataset covers from the 21 st February 2020, the beginning of localised lockdown in Italy (region of Lombardy), to the 26 th April 2020 and the following anthropogenic source categories: energy industry, manufacturing industry, road transport and aviation (landing and take-off cycle, LTO). 85 To assure easy adoption of the emission reduction factors they are produced in a format consistent with the CAMS-REG-AP emission inventory developed under the Copernicus Global and Regional emissions service (CAMS_81) (Kuenen et al., 2014;Granier et al., 2019), whose main objective is to provide gridded distributions of global and European emissions in direct support of the Copernicus Atmosphere Monitoring Service (CAMS) production chains (Marécal et al., 2015;Huijnen et al., 90 2019; Rémy et al., 2019). In the framework of CAMS, the CAMS-REG-AP emission inventory is currently used by several modelling services, mainly to provide short term air quality forecasts, long-term air quality re-analysis or policy support products. To illustrate the potential application of our reduction factors, we also performed air quality simulations to quantify and evaluate the observed changes in NO2 concentrations across Europe. We considered three emission scenarios: (i) a first one with business as usual emissions using the default CAMS-REG-AP inventory, (ii) a second one considering only the 95 traffic-related emission reductions, and (iii) a third one including the reductions from all the aforementioned sectors. The difference between scenarios allows quantifying the impact of the lockdown measures on emissions and air quality levels and, particularly, the contribution of the road transport activity to the overall reductions. The study period of these modelling https://doi.org/10.5194/acp-2020-686 Preprint. Discussion started: 22 July 2020 c Author(s) 2020. CC BY 4.0 License. exercises covers one month prior to the first day of lockdown in Italy (20 January to 20 February) and more than two months of COVID-19 lockdown conditions (21 February to 26 April). Therefore, the focus of the work is on the transition to full 100 lockdown conditions. The process toward normal conditions is a still ongoing process and will be assessed in future works.
Section 2 describes the methods and datasets used to estimate the European emission reduction factors for each one of the aforementioned pollutant sectors. Section 3 describes the setup of the modelling experiment to test the performance of the reduction factors on modelling the decrease of emissions and NO2 concentrations across Europe. Section 4 discusses the results 105 obtained in terms of emissions and NO2 level reductions. Section 5 includes our main conclusions and perspectives for future updates.

Time-, country-and sector-resolved emission reduction factors
We computed a set of emission reduction factors for Europe that vary per day, country and sector. The resulting dataset follows the sector classification reported by the CAMS-REG_AP emission inventory, which corresponds to the Gridded aggregated 110 Nomenclature For Reporting (GNFR). We considered four GNFR sectors, GNFR_A (energy industry), GNFR_B (manufacturing industry), GNFR_F (road transport) and GNFR_ H (aviation), which we assumed to be the ones suffering the largest reduction in their activity during the COVID-19 lockdowns, in line with Le Quéré et al. (2020). Other sectors potentially affected by the COVID-19 lockdown such as GNFR_C (other stationary combustion activities) or GNFR_G (shipping) were not included in this first assessment and will be addressed in future releases of the dataset. 115 In terms of spatial coverage, we included as many countries as possible that are covered by the CAMS-REG_AP European working domain (30° W -60° E and 30° N -72°N) (a complete list of the countries can be found in Granier et al. (2019)), giving a special priority to EU-30 (EU-28 plus Norway and Switzerland). A list of the countries included for each sector is summarised in Table 2. The time span of the reduction factors is from 21 February to 26 April 2020. The beginning of the 120 period corresponds to the date of the first localised lockdown in the region of Lombardy, Italy. Three distinct phases can be identified from the OxCGRT stringency index trends in Fig. 1: (i) a first phase without restrictions, with the exception of Italy (1st January to 12th March), (ii) a second phase with increasingly severe restrictions (12 to 23 March) and (iii) a third and final phase when the restrictions were at their maximum and remained almost unchanged for five weeks (23 March to 26 April).

125
We collected and processed daily measured time-series representing the main activities of each sector. We then combined this information with specific methods in order to derive daily emission reduction factors as a function of the country and sector. Table 1 summarises the main sources of information used and the countries included for each sector. The following subsections describe the data and methods for each sector along with the underlying assumptions. https://doi.org/10.5194/acp-2020-686 Preprint. Discussion started: 22 July 2020 c Author(s) 2020. CC BY 4.0 License.

Energy industry 130
We assumed the changes in emissions from the energy industry (which includes power and heat plants) to follow the changes observed in the electricity demand data reported by the European Network of Transmission System Operators for Electricity (ENTSO-E) transparency platform (Hirth et al., 2018;ENTSO-E, 2020). ENTSO-E centralizes the collection and publication of the electricity generation for each European Member State. For each country, we collected daily electricity demand data for years 2015 to 2020 (January to April). Data gaps and inconsistencies found in the original dataset were corrected using the 135 electricity generation statistics reported by the national Transmission System Operators (TSOs). For Russia, we derived the electricity demand data directly from Russia's Federal Grid Company of Unified Energy System (FGC UES, 2020).
In addition to its characteristic weekly variability, with higher values during weekdays, part of the electricity demand is driven by temperature fluctuations. Therefore, to calculate the reduction in electricity demand during the COVID-19 lockdowns, we 140 first estimated the business-as-usual (BAU) electricity demand, i.e., the demand that would have occurred in the absence of lockdowns under the same meteorological conditions. To estimate the BAU electricity demand we used ML models trained with meteorological data and other time features. This approach has been used to weather-normalize NO2 surface concentration time-series, whose variability is also partly driven by the meteorological conditions, to quantify actual reductions of NO2 during the COVID-19 lockdown (Petetin et al., 2020). More specifically, we used gradient boosting machine (GBM) models 145 trained and tuned independently for each country using daily data from January to April between 2015 and 2019. As inputs, we considered the following features: country-level daily population-weighted Heating Degree Days, date index (number of days since 2015/01/01), Julian date, day of week and a Boolean feature indicating the country-specific bank holidays. The HDD is defined relative to a threshold temperature ( ! ) above which a building needs no heating and is used to approximate the daily energy demand for heating a building (Quayle and Diaz, 1980). In order to provide a more realistic estimate of the 150 potential electricity demand for space heating on a national level, we computed country-specific population-weighted HDD values ( _ ( )) following Eq. (1): Julian day and day of week serve here as proxies for the (climatological) main drivers of the seasonal and weekly variability of the power demand, and the date index acts as the trend term. We replicated the tuning strategy previously used in Petetin et 165 al. (2020) with random search in the hyper-parameter space and rolling-origin cross-validation (appropriate for time series).
While the training and tuning of the GBM models was performed from 2015 to 2019, we used the two first months of 2020 (January-February) to test the performance of the models. Figure 2 summarizes the main statistics (normalized mean bias, NMB; normalized root mean square error, NRMSE and 170 correlation, r) obtained from the comparison between measured and ML-based electricity demand during the first two months of 2020 for selected countries. Generally, a high correlation (above 0.9) and low NMB and NRMSE (below 5%) are observed for all cases, especially in those countries with stronger lockdown restrictions such as Italy, France or Spain. The poorest performance was obtained in Finland (r = 0.33), due to a strong negative anomaly (-12% on average) of electricity demand in January-February 2020 compared to previous years used for training. Compared to most other countries, a larger NRMSE and 175 lower correlation was also found in Luxembourg. In addition, despite relatively good statistics in early 2020, the electricity demand computed in Denmark and Norway shows a substantial and unexpected increase during the COVID-19 lockdown (up to +12%). As a precautionary measure, we assumed a null reduction of the electricity demand in Denmark, Finland and Norway, and a fixed -16% reduction in Luxembourg starting the first day of the national lockdown implementation (15th of March), following the results reported by Le Quéré et al. (2020). 180 The electricity demand started to decrease by the end of February and the beginning of March 2020 compared to the BAU electricity demand estimated from the GBM models in countries where strong restrictions had been implemented. We attributed these discrepancies to the direct effect of lockdown measures, regardless of the meteorological conditions, and used them to derive quantitative daily emission reduction factors for the energy industry sector (Eq. 2) 185  where traffic activity reductions happened, followed by Spain, France, Germany, UK and Sweden. This is in line with the starting dates of lockdown restrictions in each country (Sect. 2). For Spain, reduction increased between March 30th and April 9 th , the most restrictive phase of the Spanish lockdown when only essential activities including food trade, pharmacy, and some industries were authorized. In the case of Sweden, positive values are observed for certain days until the beginning of April.
These results agree with the ones reported in Le Quéré et al. (2020), who obtained a 4% increase during the lockdown for this 200 country. It is likely that electricity demand from public and commercial services remained unperturbed as, in contrast to most countries, there was no enforced lockdown in Sweden. We also hypothesize that a voluntary self-isolation of a fraction of the population may have increased household electricity consumption. During the strictest period of the COVID-19 lockdown (23 March -26 April), Italy was the country experiencing the largest reductions (-21%), followed by Spain (-15%) and France (-

14.4%). 205
The countries for which daily reduction factors could be computed are shown in Table 1. For countries with no data, we constructed a set of reduction factors based on the average data of all the available countries except Italy, where the lockdown restrictions began approximately 3+ weeks before other countries.

Manufacturing industry 210
The reduction factors for manufacturing industry are based on the daily electricity demand reduction factors described in Sect.
2.1. We attributed 25% of the total electricity demand reduction to the reduction in manufacturing industry activity. We estimated this value considering that: (i) the European industry sector consumes 22.3% of the total final electricity demand (Eurostat, 2020a) and (ii) most of the electricity reduction during the lockdown can be linked to commercial and public services. Indeed, the manufacturing industry sector has maintained certain activities during the COVID-19 pandemic, in 215 contrast to the commercial and public services sectors that were forced to reduce or even completely halt their activities (e.g. restaurants and hotels, office buildings). Figure 4 shows, on the one hand, the evolution of the Industrial Production Index (IPI) for selected industrial branches in Spain between January 2019 and April 2020 (INE, 2020) and, on the other hand, the contribution of each Spanish commercial and public service branch to the total electricity consumption (IDAE, 2018). While certain industrial branches have suffered important decreases on their production levels during March and April 2020 (i.e. 220 production of mineral products, steel industry), the essential ones kept about the same level of productivity (i.e. pharmaceutical preparations, manufacturing of soap and detergents, petroleum refining). In contrast, office and commercial buildings, schools, universities, restaurants and hotels, which represent more than 70% of the total electricity consumption, were obliged, in most cases, to close their facilities during the lockdown.

225
The reduction of power demand attributable to the manufacturing industry sector was then translated into a total reduction in industrial activity using the national energy balances reported in Eurostat (2020a)   electricity consumption) obtained for the energy industry sector ( Fig. 3.b) were replaced by zeros for the calculations, as we consider unlikely that average increases in manufacturing industrial emissions occurred during the lockdown. In general, the trends observed in all countries follow the same pattern as the ones presented for the energy industry. During the strictest period of the COVID-19 lockdown, computed reductions are between -13 and -10% for Italy, Spain, France and UK, -4% for Germany and -0.8% for Sweden. 240

Road transport
The emission reduction factors considered for the road transport sector are based on the Google COVID-19 Community Mobility Reports (Google LLC, 2020). The Google dataset reports daily movement trends over time by geography (country and region) across different categories of places (i.e. groceries and pharmacies, parks, transit stations, retail and recreation, residential and workplaces) based on aggregated and anonymized sets of data from users who have turned on the Location 245 History setting for their Google Account on their mobile devices. For the present study, we used the mobility trends reported for the transit stations category, which includes places like public transport hubs such as subway, bus, and train stations. The assumption behind this choice is that movement trends observed in public transport hot-spots correlate with private transport trends. Reductions for each day are calculated by Google from a baseline taken as the median value, for the corresponding day of the week, over a 5-week period prior to the lockdowns (3 January to 6 February). 250 We evaluated the Google movement trends with actual measured traffic counts from the city of Barcelona (ATM, personal communication) and other major interurban roads in Spain (DGT, 2020), the latter discriminated by vehicle type (light-and heavy-duty) (Fig. 5). Note that for the Barcelona and DGT data, the information is available from 3 and 9 March onwards, respectively. In general terms, Google data reproduce the measured-based trends obtained for the city of Barcelona (BCN) and 255 the Spanish interurban roads (DGT-all), with correlations of 0.96 and 0.92, respectively. Overall, the average reductions reported by each of these three datasets are similar: -74.6% (Google), -69.1% (BCN) and -63.62% (DGT-all). Using Google data at transit stations tends to slightly overestimate the reductions observed during the weekdays. However large discrepancies are shown when comparing the Google trend against the one reported by DGT for heavy-duty vehicles (DGT-heavy). The data from the DGT reports an average reduction of heavy-duty vehicles of only -31% (more than 2 times lower than the one reported by Google), as these vehicles supported the delivery of essential goods and products (e.g. food, medical supplies). Nevertheless, we omitted the distinction between light and heavy-duty vehicle when developing the reduction factors because CAMS-REG_AP/GHG traffic-related emissions are not discriminated by type of vehicle. Consequently, our factors for the traffic sector may overestimate the overall reduction of emissions, especially in areas with a higher share of heavy-duty vehicles, typically interurban roads, and for pollutants such as PM that are emitted in a higher proportion from those vehicle categories. 265 This approach may be improved in the future but was constrained in this study by data availability. The list of countries included for this sector is summarised in Table 1. For countries without available data we constructed a set of average reduction factors considering all countries except Italy.

Aviation
We derived the reduction factors related to air traffic emissions during Landing and Take-Off cycles (LTO) in airports from 280 statistics provided by FlightRadar24 (FlighRadar24, 2020), which reports every day the total number of tracked operations per airport over the preceding 30 days. For each country, we selected the largest airport to represent a national proxy. We computed country specific daily flight operation trends using as a baseline value the average number of operations per airport from the previous year reported by Eurostat statistics (Eurostat, 2020b).

285
We started collecting the information from FlightRadar24 for all airports on 6 March, and the information from previous dates could not be retrieved as it is not archived. Therefore, our reduction factors have as initial date the 6 March in all cases, independently of the lockdown calendars. As shown in Fig. 3.d for most countries the reductions in flight activity were starting to occur during those dates and therefore the trends presented are consistent. However, in some other countries such as Italy, reductions were already in a more advanced state (first day of reduction is -15%). We do not expect this lack of information 290 to affect significantly the emission and air quality modelling results, as the contribution of this pollutant sector to total European emissions is very low, i.e. 1.1% and 0.14% to total NOx and PM10 emissions, according to the last available EMEP official reported emission data (EMEP/CEIP, 2020). We expect to complement this information from alternative sources of data in a future release of the dataset. Regarding the obtained results, it is observed that in almost all countries, the reduction levels reached values of -90% or more before the beginning of April. In contrast to road transport, there were no signs of recovery 295 during the last week of April for this sector, as the movements between countries were still restricted at that time.

Evaluating the reduction factors with air quality modelling
We performed an emission and air quality modelling study as a first demonstration and evaluation of the applicability of the developed emission reduction factors. We used the Multiscale Online Nonhydrostatic AtmospheRe CHemistry model (MONARCH) (see section 3.1) and the High-Elective Resolution Modelling Emission System version 3 (HERMESv3) (Sect. 300 3.2) both developed at the Barcelona Supercomputing Center. The simulation period for the case study is from 20 January to 26 April 2020. The study period covers one month of pre-COVID lockdown conditions (the first localised lockdowns in Europe began on 21 February in the region of Lombardy) and more than two months of lockdown conditions, including five weeks (23 March to 26 April) during which the most severe restrictions were already implemented in most (22) European countries. Therefore, the selected period of study allows analysing the changes in concentrations between the lockdown period and before. 305 Three air quality simulations were run: (i) using the default CAMS-REG-APv3.1 emissions without considering any emission reduction, hereafter referred to as baseline scenario, (ii) considering the traffic-related emission reduction factors only, hereafter referred to as covid19_traffic scenario, and (iii) including the reduction factors from the traffic, energy and manufacturing industry and aviation sectors, hereafter referred to as covid19_all scenario. We also compared the model results The rate constants were updated based on evaluations from Atkinson et al. (2004) and Sander et al. (2006). The photolysis 325 scheme used is the Fast-J scheme (Wild et al. 2000). It is coupled with physics of each model layer (e.g., aerosols, clouds, absorbers as ozone) and it considers grid-scale clouds from the atmospheric driver. The Fast-J scheme has been updated with CB05 photolytic reactions. The quantum yields and cross section for the CB05 photolysis reactions have been revised and updated following the recommendations of Atkinson et al. (2004) and Sander et al. (2006). The aerosol module in MONARCH describes the lifecycle of dust, sea-salt, black carbon, organic matter (both primary and secondary), sulfate and nitrate aerosols. 330 While a sectional approach is used for dust and sea-salt, a bulk description of the other aerosol species is adopted. A simplified gas-aqueous-aerosol mechanism has been introduced in the module to account for the sulfur chemistry, the production of secondary nitrate -ammonium aerosol is solved using the thermodynamic equilibrium model EQSAM, and a two-product scheme is used for the formation of secondary organic aerosols from biogenic gas-phase precursors. Meteorology driven emissions are computed within MONARCH. Mineral dust emissions are calculated with an updated version of Pérez et al. 335 (2011) scheme, the sea salt aerosol emissions following Jaeglé et al. (2011), and biogenic gas-phase species using the MEGANv2.04 model (Guenther et al., 2006). The model provides operational regional mineral dust forecasts for the World Meteorological Organization (WMO; https://dust.aemet.es/), and participates to the WMO Sand and Dust Storm Warning Advisory and Assessment System for Northern Africa-Middle East-Europe (http://sds-was.aemet.es/). Since 2012, the system contributes with global aerosol forecast to the multi model ensemble of ICAP initiative (Xian et al., 2019) and since 2019, it 340 is a candidate model of the CAMS -Air Quality Regional Production (Marecal et al., 2015).
In this work, the model is configured for a regional domain covering Europe and part of northern Africa. The rotated lat-lon projection is used, with a regular horizontal grid spacing of 0.2 degrees, and the top of the atmosphere is set at 50 hPa using 48 vertical layers. Figure S1 displays the domain of study. Meteorological initial and boundary conditions were obtained from 345 the ECMWF global model forecasts at 0.125 degrees and chemical boundary conditions from the CAMS global model forecasts at 0.4 degrees (Flemming et al., 2015). For an efficient execution of the modelling chain, the autosubmit workflow manager is used (Manubens-Gil et al., 2016).

HERMESv3 emission system
The original annual CAMS-REG-APv3.1 emission inventory was processed using the HERMESv3 system, an open source, 350 stand-alone multi-scale atmospheric emission modelling framework developed at the BSC that computes gaseous and aerosol emissions for use in atmospheric chemistry models ). The HERMESv3 system was used to remap the original CAMS-REG-AP data (0.1x0.05 degrees) onto the MONARCH modelling domain and to derive hourly and speciated emissions. Aggregated annual emissions were broken down into hourly resolution using the emission temporal profiles reported by Denier van der Gon et al. (2011). The speciation of NMVOC and PM emissions was performed using the split 355 factors reported by TNO (Kuenen et al., 2014).
For the covid19_traffic and covid19_all scenarios, the estimated reduction factors (Fig. 3.a,b,c,d) Figure 3.e illustrates the COVID-19 daily temporal factors for the road transport sector in selected countries. The original daily profile for this sector, which is used in baseline scenario, is also plotted for comparison purposes.
In general, the temporal disaggregation of emissions would require the sum of the daily weight factors to be 366 (as in this 370 case the year of study is a leap year). Nevertheless, and due to the application of the reduction factors, the sum of the COVID-19 daily factors do not add up to this number, which allows simulating time-resolved emission reductions.

Observational dataset
The GHOST project is a BSC initiative dedicated to the harmonisation of publicly available global surface observations (most notably air quality pollutants) and metadata, for the purpose of facilitating a greater quality of observational/model comparison 375 in the atmospheric chemistry community (Bowdalo, in preparation). Numerous networks are currently processed and contained under the umbrella of GHOST including, among other, the EBAS and EEA networks. For each network, all relevant numerical and textual metadata (e.g. station classifications, measurement methodologies) is standardised and all data is passed through numerous quality control tests, giving detailed quality assurance (QA) flags.

380
In this work, we used the NO2 near-real time EEA data. We selected rural and urban background stations located at selected countries (Italy, Spain, France, Germany, UK and Sweden). In the case of urban background stations, we selected those located in Milano, Madrid, Paris, Berlin and London. For Sweden, and due to the low density of stations found in individual cities (e.g. Stockholm, 1 station), we decided to consider all urban background stations available country wise (6). GHOST provides a wide range of harmonized metadata and quality assurance (QA) flags for all pollutant measurements. In this study, we took 385 benefit of these flags to apply an exhaustive QA screening. More details on the QA flags used can be found in Appendix A.
Note that for Italy, there is a data gap between 1 February and 13 February in all stations. We nevertheless decided to keep this country in our evaluation study since it is one of the European countries most affected by the COVID-19 pandemic and the data gap does not affect the lockdown period. In the case of Sweden, only 1 rural background station was available for the entire country, which may reduce the representativity of the computed results. A detailed description of the stations is available 390 in Table S1 and Fig. S1 of the supplementary material.

4
Results and discussion and third week of March, when several European countries enforced national lockdown restrictions. After this period, there was a stabilization of the emission reductions until approximately the 19 April. Thereafter, a slight recovery of the emission levels started to occur, which is consistent with the recovery of traffic activity shown in Fig. 3.c. Overall, and when comparing the baseline and covid19_all scenarios, the reduction of total emissions is -33% for NOx, -8% for NMVOC, -7% for SOx and -7% for PM2.5. The contribution of the traffic sector to total reductions is especially relevant for NOx (90%), NMVOC (87%) 410 and PM2.5 (82%) while for SOx most of the total reduction can be attributable to the decreases in the energy and manufacturing industries (97%), according to the results shown by the covid19_traffic scenario. Figure 7 (e, f) illustrates the average and 5th/95 th percentiles (p05, p95) of the daily relative changes [%] in the gridded NOx emissions for Italy and Sweden. The results were computed considering all the grid cells within each of the countries. In Italy, the last two weeks of March and first two weeks of April certain shows areas of the country reaching reductions up to -75%, whereas in other areas less affected by 415 anthropogenic (and particularly road transport) emissions the reductions were significantly lower (~ -25%). In the case of Sweden, the reductions ranged between -6% (p95) and -36% (p05). Germany and Sweden present stronger average reductions than the ones reported at the EU-30 level (-33% and -7%, respectively), and Italy and France are the two countries with the largest reductions (-50% for NOx and -12% for SOx). For NOx, minimum and maximum daily emission reductions are in general relatively close to the average (e.g. Italy: avg = -50%, min = -47% and max = -56%; Spain: avg = -40%, min = -43% and max = -46%). In contrast, there are large differences among 425 the average, minimum and maximum daily SOx changes, especially in Germany (Sweden) where changes in emissions go from 0.6% (0.15%) to -12% (-5%). The different behaviours observed for NOx and SOx are related to the different trends of the road transport and energy industry (Fig. 3.a and c). The daily variability of the reduction factors for road transport is generally low; in the case of the energy industry large day-to-day variations are observed.

430
Despite having experienced one of the largest reductions in road transport activity (more than -80%), Spain was the country with the lowest decrease in total PM2.5 emissions (-4.3%), and the second lowest in terms of NMVOC (-4.4%). Sweden shows a PM2.5 emission reduction of -7.6%, almost two times larger than Spain and very close to Italy (-9.2%), despite its lower traffic activity decrease (less than -40%). This is explained by the different contributions of the road transport sector contribution to total emissions in each country. Figure 9 shows the relationship between the reduction of traffic activity and 435 contribution of the road transport sector to total emissions per country and pollutant. In the case of Sweden, road transport represents around 21.3% of total PM2.5 emissions, while in Spain the contribution is just 7.9%. Similarly, in the case of NMVOC emissions the contribution of road transport emissions is 15.9% in Italy and 8.9% in France, while in Spain is only 4.3%.  Fig. 11. In both cases, the results are presented separately for each of the emission scenarios considered: baseline (in magenta), covid19_traffic (in green) and 445 covid19_all (in blue). Statistical parameters computed on an hourly basis (i.e. mean bias, MB; root mean square error, RMSE; correlation coefficient, r) are presented for each emission scenario, country and station type for the pre-lockdown (20 January to 20 February) and most restrictive lockdown period (23 March to 26 April) (Fig. 12). (For the pre-lockdown period, the calculated statistics are equal for all scenarios, as no emission reductions are considered during that time.) The computation of statistics during the pre-lockdown period allows quantifying the performance of the system under BAU conditions. Table 2  450 summarises the absolute and relative changes of NO2 concentrations at each station type and country between 23 March and The MONARCH model is capable of reproducing fairly well the urban background NO2 observations during the pre-lockdown period, particularly in London (MB = -0.25 µg·m -3 , RMSE = 16 µg·m -3 , r = 0.74), Madrid (MB = -4 µg·m -3 , RMSE = 19 µg·m -455 3 , r = 0.64) and Paris (MB = -7.7 µg·m -3 , RMSE = 13 µg·m -3 , r = 0.78). Milano is the location with the largest MB (-14 µg·m -3 ) and RMSE (22 µg·m -3 ). The relatively low performance in Milano may be related with the inability of reproducing the strong atmospheric stability conditions of the Po Valley region, a general problem for chemical transport models. After the implementation of the national lockdowns, a decrease in NO2 is simulated in all sites for both the covid19_traffic and covid19_all scenarios. Nevertheless, the decreasing rate strongly varies from one country to the next. In Madrid and Paris, 460 NO2 concentrations drop abruptly just a few days after the beginning of the lockdown, while in Milano, Berlin and London the decreases occur at a slower pace. These results are consistent with the traffic activity reduction trends computed for these countries (Fig. 3.c). The statistics computed for the most restrictive lockdown period (23 March to 26 April) clearly reveal a general improvement of the model performance when the emission reductions are considered. As shown in Fig. 12, the calculated MB and RMSE values for the baseline scenario are significantly reduced when considering the covid19_traffic and 465 covid19_all scenarios, especially in Madrid, Paris and London where overestimations of 9 to 14 µg·m -3 are drastically reduced to 1 to -1.5 µg·m -3 . In Berlin the performance of the model slightly decreases when considering the lockdown scenarios. Both the MB and RMSE of the baseline scenario remain lower in magnitude. This feature is attributed to a significant increase in observed NO2 during the week of 7 April that neither the baseline nor the covid scenarios capture, either due to missed emission activity changes or errors in meteorology. In terms of correlation, no significant changes are observed when comparing the 470 baseline and covid scenarios (for all cases except Milano values stay above 0.6). The computed absolute and relative decreases of NO2 urban background levels reveals that the differences between the covid19_traffic and covid19_all scenarios are generally low, i.e. the decrease in modelled NO2 concentrations is mainly driven by reduction of road traffic emissions. This is consistent with the large contribution of the traffic sector to total NOx emission reductions as discussed in Sect. 4.1.
When it comes to rural background levels, pre-lockdown statistics also indicate a good capability of MONARCH in 480 reproducing observed values, particularly in France (MB = 1.5 µg·m -3 , RMSE = 2.8 µg·m -3 , r = 0.93) and Spain (MB = -0.16 µg·m -3 , RMSE = 1.3 µg·m -3 , r = 0.52). A persistent overestimation is observed in Germany, UK and Sweden (MB between 3.5 and 4.6 µg·m -3 ), while in Italy the system tends to underestimate (MB = -3.6 µg·m -3 ). The overestimation in Germany, UK and Sweden occurs mainly at night-time (not shown). Similar to what is observed at urban background sites, modelled and observed concentrations between 23 March and 26 April tend to be more in agreement when considering the emission reduction 485 scenarios. The UK and Germany are the countries were the performance improves more, with MB values going from 7.5 and 2.3 µg·m -3 (baseline) to 3.4 and 0.42 µg·m -3 (covid19_traffic) and 3 and 0.29 µg·m -3 (covid19_all). On the other hand, the improvement is not obvious in Italy, as the model shows a negative bias during the pre-lockdown period and the lockdown scenarios constitutes an important reduction of the modelled values. However, the trend is in agreement with results in Spain, France and Germany but with some additional underestimations. The rural background NO2 concentrations in the two 490 lockdown scenarios are substantially lower than in the baseline run. Nevertheless, the relative decreases are generally lower than in urban environments. France (-44% and -42% for covid19_all and covid19_traffic, respectively) and Italy (-43% and -41%) are the countries that experience the largest decreases, followed by Spain, UK and Germany (around -30% and -28% in all of them). In Sweden, relative reductions are almost equal to the ones obtained in urban background locations (-12% and -11%). Although no robust conclusions can be extrapolated as the results are based on only one rural station, the similar 495 reductions obtained in both environments could be related to the soft restrictions implemented in this country. When comparing the covid19_all and covid19_traffic scenarios, only around 4 to 8% of the total reduction can be attributed to non-traffic sources.

Conclusions
This paper presents a dataset of daily, sector-and country-dependent emission reduction factors that allows quantifying the 500 impact of the COVID-19 lockdown on European primary emissions and air quality levels. The reduction factors are provided for a period that goes from 21 February, when the first European localised lockdown was implemented in the region of Lombardy (Italy), to 26 April 2020, and for the four emission sectors presumably most affected by the mobility restrictions, i.e., road transport, energy industry, manufacturing industry and aviation. Our emission reduction factors are based on a wide range of information sources, including open access and near-real time measured activity data, proxy indicators and other 505 available reports. We also make use of machine learning techniques trained with meteorological data to estimate reductions in electricity consumption.
We combined the computed reduction factors with the Copernicus CAMS European emission inventory using adjusted temporal profiles in order to derive time-resolved emission reductions per country and pollutant sector. We also performed an 510 air quality modelling study to evaluate the potential of the computed emission reductions on reproducing observed NO2 concentration decreases in selected rural and urban background regions across Europe (Italy, Spain, France, Germany, UK and • During the most severe lockdown period (23 March to 26 April), estimated emission reductions at the EU-30 level 520 were -33% for NOx, -8% for NMVOC, -7% for SOx and -7% for PM2.5, with road transport being the main contributor to total reductions in all cases (85% or more) except for SOx, for which reductions were mainly driven by the energy and manufacturing industry sectors.
• Italy, France and Spain are the countries that experienced the major NOx and SOx emission reductions (up to -50% and -12%, respectively), a result that is in line with the strong lockdown restrictions implemented by their 525 corresponding governments. On the contrary, Sweden shows reductions of only -15% (NOx) and -2.5% (SOx) due to implementation of national recommendations instead of a state-enforced lockdown.
• Despite showing lower reductions of road transport activity, calculated reductions of total PM2.5 in Sweden are much larger (-8%) than in Span (-4%). This is due to the variation in the contribution of the road transport sector to total emissions from country to country. While in Sweden road transport represents around 21.3% of total PM2.5 emissions,530 in Spain this contribution is of just 7.9%. A similar outcome is obtained for NMVOC when comparing traffic activity and total emission reductions in Spain and France.
• According to air quality modelling results, the larger decreases of urban background NO2 levels occured in Madrid (-58%) and Milano (-56%). The calculated NO2 relative reductions at rural background areas are generally lower, with France (-44%) and Italy (-43%) being the countries that experience the largest decreases. 535 • In both urban and rural environments, the comparison between covid19_traffic and covid19_all results, indicates that the road transport sector is on average responsible for 90% of the total NO2 reductions, with the largest and lowest contributions found in Milano (97%) and Berlin (76%), respectively.
• Overall, we found the performance of the modelled NO2 results to clearly improve when considering the emission reduction scenarios. Calculated MB values for the covid19_traffic and covid19_all scenarios are significantly lower 540 than the ones estimated for the baseline scenario, especially in Madrid, Paris and London where overestimations of 9 to 14 µg·m -3 are drastically reduced to 1 to -1.5 µg·m -3 . On the other hand, the improvement is not so obvious at locations where the modelled results already display an important bias during the pre-lockdown period.
In this work we present and evaluate a methodology not only to calculate time-resolved emission reductions associated to the 545 COVID-19 lockdown, but also to adapt them for air quality modelling purposes, which may be relevant for the modelling community. There are, however, some limitations associated to the current version of the reduction factors dataset. First, and most importantly, emission changes in each sector were inferred from changes observed not directly in emissions but in general activity proxies such as electricity demand or traffic indicators. The use of such general indicators may lead to disregard changes associated to specific processes or sources. For instance, and as discussed in this work, comparisons against observed 550 traffic counts showed that the Google movement trends are not representative of observed changes in heavy-duty vehicle's activity, and that their use may lead to a potential overestimation of the overall traffic activity reduction, especially in interurban roads, where the share of these vehicle categories is more important. In the case of energy industry, the association between changes in electricity demand and emissions from power and heat plants neglects potential changes in the national power mixes. As recently presented by the International Energy Agency (IEA), certain countries have shifted their electricity 555 production towards renewables following lockdown measures due to low operating costs and priority access to the grid through regulations, among other (IEA, 2020). Omitting this aspect may be leading to an underestimation of the emission reductions for this sector, and therefore will be revised in future versions of the dataset. Finally, in the manufacturing industry sector the same reduction factors are assumed for all the industry branches. Yet, information reported by national industrial production indexes are indicating that not all industrial sectors were affected in the same way by the lockdown restrictions. For example, 560 Spanish pharmaceutical industries experienced no changes in their activity during March and April, while industries related to the production of mineral products showed significant decreases. Regarding this last point, it is important to note that the specificity of the computed reduction factors also depends upon the degree of sectoral disaggregation used to report the original CAMS inventory. In the case of the manufacturing industry sector, all emissions are reported under a unique category, which hampers the consideration of industrial divisions. One last important shortcoming is related to the spatial variability of the 565 proposed reduction factors. In its current version, the reduction factors are country-dependent and therefore do not take into account potential variations within each country. This includes, for instance, the contrast between the large cut in road traffic to and from airports on the one hand and the traffic congestion of heavy-duty vehicles at the national borders captured by the Copernicus satellite images on the other (EU, 2020). This aspect will be also relevant when extending the time series of the dataset and including the period when governments started to soften lockdown measures. In some countries such as Spain this 570 process was implemented heterogeneously across the different administration units.
Despite the aforementioned limitations, we believe that providing these timely emission modelling results will help with the understanding of air quality related aspects of the pandemic and also to better prepare in case of new waves or resurgences.
As a matter of fact, this dataset supports a number of studies that are on-going in particular within CAMS and under the Global 575 Atmosphere Watch Programme of the World Meteorological Organization (WMO/GAW). Future works will focus on amending the shortcomings mentioned above, extending the number of sectors considered, in particular the residential/commercial and shipping sectors, and covering the transition period towards the post-lockdown conditions. The investigation of the calculated emission reductions obtained when combining the reduction factors with the new CAMS emission temporal profiles (Guevara et al., submitted) will be also studied. New datasets and information sources will become 580 soon available and therefore allow for an improvement of the representativeness of the current emission reductions. Moreover, the evaluation of the reduction factors in reproducing observed changes in other air pollutants such as O3 or PM2.5 will be also addressed in the future. We also expect to perform inter-comparisons of our modelled results against reductions associated to the COVID-19 lockdown derived from satellite-based observations. Using the information provided by GHOST (Globally Harmonised Observational Surface Treatment), we applied numerous QA screening to the NO2 dataset, in order to remove : missing measurements (flag 0), infinite values (flag 1), negative measurements (flag 2), zero measurements (flag 4), measurements associated with data quality flags given by the data provider which have been decreed by the GHOST project architects to suggest the measurements are associated with substantial 590 uncertainty or bias (flag 6), measurements for which no valid data remains to average in temporal window after screening by key QA flags (flag 8), measurements showing persistently recurring values (rolling 7 out of 9 data points; flag 10), concentrations greater than a scientifically feasible limit (above 5000 ppbv) (flag 12), measurements detected as distributional outliers using adjusted boxplot analysis (flag 13), measurements manually flagged as too extreme (flag 14), data with too coarse reported measurement resolution (above 1.0 ppbv) (flag 17), data with too coarse empirically derived measurement 595 resolution (above 1.0 ppbv) (flag 18), measurements below the reported lower limit of detection (flag 22), measurements above the reported upper limit of detection (flag 25), measurements with inappropriate primary sampling for preparing NO2 for subsequent measurement (flag 40), measurements with inappropriate sample preparation for preparing NO2 for subsequent measurement (flag 41) and measurements with erroneous measurement methodology (flag 42).

6
Data availability 600 The computed emission reduction factors per country, sector and day are provided in the supplementary material.