Effect of accounting for public holidays on skills of atmospheric composition model SILAM v.5.7

Changes in anthropogenic activity during public holidays influence air pollutant concentrations. The objective of this study is to quantify the public holiday’s effect on air quality and to analyse the added value of accounting for the holidays in AQ modelling and forecasting. Spatial and temporal distributions of atmospheric concentrations of the major air pollutants (PM2.5, PM10, SO2, CO, NO2, NOX, and O3) were considered at the European scale for all public holidays of 2018. Particular 10 attention was given to the events with the most-pronounced continental or regional impact: Christmas and New Year, Easter, May vacations and last days of Ramadan. The simulations were performed with the Eulerian chemistry transport model SILAM v.5.7. Three model runs were performed: the baseline with no treatment of holidays, the run considering holidays as Sundays, and the run forcing 80% reduction of emissions during holidays, for the week-day sensitive sectors. The emission scaling was 15 applied on a country basis. The model predictions were compared with in-situ observations collected by the European Environment Agency. The experiment showed that even conservative treatment of official holidays has a large positive impact on NOx (up to 30% of bias reduction in the holiday days) and also improves the CO, PM2.5 and O3 predictions. In many cases, the sensitivity study suggested deeper emission reduction than the level of Sundays. An individual consideration of the holiday events in 20 different countries may further improve their representation in the models: specific diurnal pattern of emissions, additional emission due to fireworks, different driving patterns, etc.


25
Air quality (AQ) and its temporal and spatial changes are determined by human activities via the release of various aerosols and gases (Derwent and Hjellbrekke, 2012;Fu et al., 2020;Hassan et al., 2013;Karl et al., 2019;Kukkonen et al., 2020;Lehtomäki et al., 2018;Shi et al., 2019), and modulated by meteorological conditions (Jacob and Winner, 2009;Jhun et al., 2015;Singh et al., 2013;Sofiev et al., 2020). 30 The ability of atmospheric composition models to follow the temporal variability of air pollution critically depends on representation of the emission temporal behavior by the inventories used by the models. Arguably the most-difficult task is to catch the variations originating from rare events. Changes https://doi.org/10.5194/gmd-2021-52 Preprint. Discussion started: 12 April 2021 c Author(s) 2021. CC BY 4.0 License.
However, the weekend and (some) holiday effects have certain similarities, which might allow drawing 40 an analogy between weekday vs. weekend and holiday vs. non-holiday pollution levels.
Majority of currently available emission inventories are built as gridded yearly or monthly totals for the key primary pollutants (Frost et al., 2013;Granier et al., 2019Granier et al., , 2011, (https://eccad.aeris-data.fr/, access 5.2.2021). Temporal variations at shorter time scales received less attention but their impact on AQ itself and the model's ability to reproduce the concentration has been considered in several studies as 45 well (Fu et al., 2013;Gioli et al., 2015;Guevara et al., 2017Guevara et al., , 2021Iriti et al., 2020;McGraw et al., 2010). Several studies have demonstrated the crucial role of spatial and temporal resolution of emission inventories in environmental science, air quality modeling, and air pollution policy making (Frost et al., 2013;Gioli et al., 2015;Zhao et al., 2015;Zhou et al., 2020).
A number of observations-based studies focused on effects of the weekends and, sometimes, specific 50 holidays on pollutants concentrations Forster and Solomon, 2003). Lonati et al. (2006) examined the weekend effect for particulate matter (PM10 and PM2.5) emissions from traffic sources in the city of Milano. The research indicated that concentrations of these compounds in the urban area were lower than the levels during the weekdays. Gour et al. (2013) considered differences in the pollution levels during weekends and weekdays in Delhi and showed that pollution variation follows 55 the pattern of working activities on weekends and weekdays. Parra and Franco (2016), pointed out that the concentration of NO2, NOX, CO, and PM2.5 in working days is higher than that at the weekend, but the concentration of O3 in working days is lower than that of the weekend, due to ozone titration. In (2017), Ding et al. reported that during the Chinese New Year the NOx emissions are usually lower by about 10% reflecting the lower business and industrial activities.. In a recent study, Hua et al. (2021) 60 estimated the holiday effect of PM2.5 and NO2 by a Generalized Additive Model (GAM) with regard to time and meteorological parameters at 34 air quality monitoring stations during the five heating seasons from 2014 to 2019 in Beijing. According to their results, the holiday effect was much stronger than the weekend effects with increasing PM2.5 by 2% to 30% and decreasing NO2 in contrast. Khalil et al. (2016) analysed hourly measurements of nitrogen oxide (NOx), non-methane hydrocarbons 65 (NMHCs), ozone (O3), sulphur dioxide (SO2), PM2.5, and PM10 collected at the coastal town of Yanbu, Saudi Arabia during weekends, Eids, Ramadan, and the Hajj periods and demonstrated that the ozone concentrations stay practically the same over these holiday days but the precursor levels are Recently, various methods based on observed data and models were applied to measure the impact of COVID-19 lockdown on air pollution. These studies investigated the role of transport and industry sectors (as the main sources of air pollution) on pollutants concentrations during the lockdown (Fan et 80 al., 2020;Grivas et al., 2020;Huang et al., 2020;Menut et al., 2020;Sharma et al., 2020;Wang and Su, 2020).
The above works showed that the effects of isolated events, such as public holidays, can be substantial.
Yet its systematic analysis at large scales (e.g., a continent and a full year) is missing and a systematic approach to their incorporation into AQ models is yet to be developed. 85 The goal of the current paper is to address this gap and to make the first step towards incorporation of the public holidays into the regular atmospheric composition and air quality modelling in Europe. We quantified the added value of a comparatively primitive and conservative inclusion of official holidays into temporal profiles of emission of air pollutants. Secondly, a sensitivity study was performed demonstrating the extent of the necessary adjustments and potential benefits of a more detailed analysis 90 of each specific holiday event.
The paper is organised as follows. The next section presents the methodology of the study: information

European Holidays 100
We collected a list of official holidays in Europe from the Calendarific global holidays API consider the events marked with "National holiday", "Local holiday" or "Common local holiday" as holidays (see examples for some European countries in Table 1 -Table 3).
The model computations included all holidays in 2018 but, for the sake of brevity, the analysis below 105 will concentrate on the Christmas and New Year weeks, Easter (analysed at the European scale), and the Festival of Breaking the Feast at last days of Ramadan (Eid al-Fitr, analysed for Turkey). 17.01.2021) is an offline 3D chemical transport model (Sofiev et al., 2015a), also used for emergency decision support (Sofiev et al., 2006) and inverse atmospheric composition problems (Sofiev, 2019;115 Vira and Sofiev, 2012). The model incorporates Eulerian and Lagrangian dispersion frameworks and a variety of chemical / physical transformation modules covering the troposphere and the stratosphere (Carslaw et al., 1995;Damski et al., 2007;Gery et al., 1989;Kouznetsov and Sofiev, 2012;Sofiev, 2002Sofiev, , 2000Sofiev et al., 2010;Yarwood et al., 2005). SILAM features a mass-conservative positivedefinite advection scheme based on principles laid down by M.Galperin (Galperin et al., 1996). The 120 model can be run at a range of resolutions and coverage starting from a kilometre scale over a limited area and up to the whole globe Kouznetsov et al., 2020;Petersen et al., 2019;Sofiev et al., 2020Sofiev et al., , 2015bXian et al., 2019). The vertical structure of the modelling domain consists of stacked layers starting from the surface. The layers can be defined either in zor hybrid sigma-pressure coordinates. The model can be driven with a variety of numerical weather prediction or climate models. 125

Simulation setup
The simulations were performed for the whole year of 2018 for the European domain with the setup following the operational configuration of SILAM in the Copernicus Atmospheric Monitoring Service (CAMS) regional air quality forecasts, as of November 2020 (https://atmosphere.copernicus.eu, access 20.02.2021). The only exception was a twice coarser grid resolution to reduce the computational costs 130 (Table 4).  (Wesely, 1989) for gases, (Kouznetsov and Sofiev, 2012)  profiles (Granier et al., 2019). In the CAMS-regional operational setup, the anthropogenic emissions are used without accounting for public holidays.
To assess the sensitivity of air concentrations of pollutants to holidays, three SILAM runs were made: 140 the baseline (hereinafter, the BL case), with the holiday days considered as Sundays (the HS case), asensitivity run with holidays getting 80% of emission reduction for the sectors affected by the DOW profile (the R3 case, considered in the Discussion section). Technically, the emissions were adjusted by altering the DOW scaling coefficients for dates and countries where the holidays occur. For the HS case the coefficients were set to their Sunday values, and for the R3 case they were forced to 0.

150
For evaluation of the simulations we used the hourly data of the AQ monitoring stations downloaded from the European Environmental Agency portal (EEA, http://discomap.eea.europa.eu/map/fme/AirQualityExport.htm, visited 10.01.2021). Since we focus on regional-scale effects, a subset of representative stations was selected, namely, the stations classified from 1 to 7 according to Joly and Peuch (2012) classification. This dataset is also used for the 155 operational CAMS-regional evaluation (751 stations over the European domain).
The effect of holidays was considered for the main pollutants observed by the EEA network: PM2.5, PM10, SO2, CO, NO2, NOx, and O3. Five statistics were considered following the CAMS evaluation standards: bias, fractional bias (FracB), Pearson correlation coefficient (corr), RMSE, and fractional gross error (FGerr). 160 We considered the effect of holidays at two temporal scales. The short-term impact was analysed for the one-week period centred around each holiday day. For each day of this period, the spatial statistics were computed across the observational stations, and evolution of these statistics from day to day was compared between the SILAM runs. The long-term longitudinal effect was analysed at annual level for the whole 2018 and attention was given to the temporal statistics computed for the stations time series. 165 Since the diurnal profile of emission during holidays is unknown, albeit probably specific for each event and country, the current study mainly used daily averaging of both observational and model data for computations of the statistics.

Short-term impact of public holidays 170
The impact of holidays on the SILAM spatial skills was the largest for the Christmas week ( Figure 2a).
As expected, the Christmas period is characterised by lower emissions, which resulted in a high bias of the BL model run and almost 50% growth of the RMSE compared to surrounding days. The reduction of emission in the HS run improved the performance but did not eliminate the problem completely.    In the Muslim countries (Turkey, Albania), the Ramadan month is not a public holiday as a whole, just 205 working hours are reduced, which is not reflected in the HS run. Only the last three days of Ramadanthe Ramadan Feastare the public holidays in Turkey (Table 3, Figure 6 for NO2, Supplementary material for other species, Figs. S19-S24). For these days, there are distinct differences between the BL and HS model runs. However, similar to Easter and the May day, the model is generally low biased for NO2 in Turkey during this period, therefore the additional reduction of the concentrations is, formally 210 speaking, not an improvement: the negative bias increases. Nevertheless, it is a step in the right direction, as seen from the reduced variations of the model skills of the HS run: handling a flat systematic bias is easier than a scatter. The NO2 under-estimation in Turkey probably originates from the understated emissions, which update would resolve the issue. Due to this under-estimation, it is difficult to estimate how conservative the Sunday-level emission reduction is for these holidays ( Figure  215 6 b, c).
Unlike the Christmas and Easter holidays, which exist in most European countries, especially those with the highest density of the observational network and the strongest emission, the Ramadan Feast days have a substantial effect only for the Turkish stations. At the European scale, the effect is negligible.

Long-term statistics
At the annual scale, the impacts of holidays on the model performance is limited. The reduction affects only the days with changed emissions and practically do not influence already the next day. The mostsignificant impact was for Christmas and New Year weeks but even for them the effect faded out by the next day. According to the annual statistics, the HS run performed slightly better than the BL: the model 225 bias and RMSE in HS run are lower and correlation is higher than in the BL run. Quantitatively, at annual level the overall effect for NO2 was less than 1%, which reflects the typical number of holiday https://doi.org/10.5194/gmd-2021-52 Preprint. Discussion started: 12 April 2021 c Author(s) 2021. CC BY 4.0 License. days in a year (< 3%) and at most 30% improvement during these days. Impact on other species was lower than that for NO2.

Holiday effect on model skills: episodically significant, noticeable at annual level
The simulations presented in the previous section confirmed that the official holidays substantially affect air quality, as also shown in the studies outlined in the Introduction. The holiday incorporation into the simulations as Sundays, being very simple technically, brings noticeable improvement of the model skills for the days with the modified emission. Since the number of such days in each year is < 235 3%, the overall improvement of the annual skills is expectedly within 1%, which is quite significant at such level of aggregation.
The suggested simple approach should be considered as only the first step. Holidays are characterised by redistribution of emission due to changing traffic structure, shift of activities from office areas to suburbs, etc. Incorporation of these effects can further improve the model skills but will require 240 quantitative information on such redistribution at the European level. Some support can be found from traffic information, which is presently not available at continental scales.

Sunday-based emission reduction for holidays is a conservative estimate
The simulations also suggested a comparatively simple way to achieve a more significant gain: the Sunday emission scaling (Figure 1) can be amplified. In a few cases, especially for the Christmas and 245 New Year, the actual emission rates might be much lower, whereas for some events the emission of some species might increase. Thus, the New Year night celebration in many countries involves fireworks, which add substantial amount of PM. The second issue is that the Sunday diurnal profile of traffic (also other sources) is substantially different from that of the weekdays. In the present version of SILAM this difference is not accounted for, which evidently limits the model performance and the gain 250 due to the holiday incorporation.
In order to estimate the actual emission reduction over the Christmas and New Year week, we performed a sensitivity simulation HolidayPlus (R3), for which the emission was reduced by 80% (see Methodology section for details). Being a clear overshot, this run was deemed as the limit-from-below of the emission during the holidays. The corresponding observed and modelled time series of NO2 255 concentrations at a station in the Netherlands are presented in Figure 7 for the Christmas week. The The present findings are consistent the estimates of observations-based studies. Thus, with Hua et al (2021) also found that the holiday effect is much stronger than the weekend effects. They noticed the 270 opposite signs for PM2.5 and NO2: average increase of about 22% and average decrease of about 11%, respectively. Similarly, Retama et al., (2019) reported substantial effect of fireworks on PM at night and the following morning of Christmas Day and the New Year's day. Along the same lines, Rozbicka and Rozbicki (2016), demonstrated that daily mean ozone concentration and maximum ozone peaks are respectively 13% and 8% higher than those on weekdays, which also indicates a reduction in NO2 275 concentrations of about 20%. Conversely, Nodehi et al. (2018) study showed that the Norooz holidays (the Iranian New Year, or a spring festival), are characterised by a reduction of concentration of PM2.5 due to the reduction of the working activities and no massive fireworks. The reported reduction of PM2.5 concentration during the Ramadan Feast holidays is quite close to our estimates.

Regional specifics 280
The impact of holidays varies from country to country with substantial differences visible even at a subcountry level. The maps of the station-wise temporal correlation coefficients for hourly NO2, CO, O3, and PM2.5 concentrations (Figure 8, Figure 9) reveal a strong inhomogeneity of the effect for Christmas and New Year weeks. The effect can dramatically vary even within a single countryas seen from the comparison of maps of Figure 8 and country-median correlation coefficient of Figure 9. 285 https://doi.org/10.5194/gmd-2021-52 Preprint. Discussion started: 12 April 2021 c Author(s) 2021. CC BY 4.0 License.
In the case of NO2, correlation increases, e.g., in Northern Germany, Italy, Poland and Eastern part of Finland for both HS and R3 runs. Conversely, there was no effect or even deterioration of skills in Southern Germany, Northern France, Madrid region, etc. Other species showed qualitatively similar patterns but lower gains and losses. Surprisingly, skills over most of France are generally worse for the HS run and much worse for R3 indicating a substantially different pattern of activities during holidays, 290 compared to those of the neighboring countries. The R3 run, which was planned as an overshot, showed strong improvement of temporal correlation in Eastern Europe, Central and Northern Italy and Northern Germany. Therefore, one can argue that the 5-fold emission reduction in these countries / regions might be not that much of an exaggeration. The issue deserves a more detailed analysis accounting for the varying traffic patterns and effects on days preceding to and following the official holidays.

Summary
Incorporation of information on public holidays in emission of the affected anthropogenic sectors leads to substantial short-term improvements of the SILAM model scores, even if done conservatively. The largest impact was found for NOx, which is controlled by the changes of the traffic intensity. Certain improvements were also found for PM2.5 and ozone but the signal was weaker than that for NOx. 310 The effect of the emission reduction during holidays may look detrimental in case of a systematic under-estimation in some regions. However, in majority of such cases the bias became more homogeneous in time and thus easier to handle with, e.g., emission corrections via data assimilation or development of new emission inventories.
The sensitivity runs confirmed that the Sunday emission level, in many cases, is a too conservative 315 proxy for the public-holiday emission. Thus, the reduction during Christmas and New Year holidays of 2018 was closer to a factor of 4 in Western Europe and possibly even stronger in Eastern Europe.
The current experiment used the prescribed sector-specific diurnal profiles of emission intensity, same for weekdays, weekends and holidays. Incorporation of specific profiles for weekends and holidays, when they become available, will further improve the quality of the model predictions. 320 The proposed method of handling emission reduction in AQ models, albeit very simple and with a room for improvement, gives noticeable gains in the model performance scores. The method is straightforward to implement in AQ models and can be considered as an easy way to significantly improve the model prediction skills for the periods of public holidays. An in-depth analysis of the specific holidays and related traditions in specific countries, such as fireworks in New Year night, 325 would, most probably, lead to further improvements of the AQ predictions.

Code and data availability
SILAM is an open-code system and can be obtained from the GitHub open repository (https://github.com/fmidev/silam-model, Kouznetsov and Delgado, 2021). The simulation results are 330 available on request from the authors of the paper.

Author contribution
The authors jointly devised the project and developed the paper concept. YF contributed to the implementation of the research and analysis of the results, and drafted the paper. RK performed the SILAM computations and contributed to the analysis. MS contributed to the analysis, drafted the 335 Discussion and contributed to other sections of the paper. All authors edited the final text.

Acknowledgements
The study was performed within the scope of Academy of Finland project GLORIA (grant Nbr 310372