ETC/ATNI Report 14/2019: Statistical modelling for long-term trends of pollutants. Use of a GAM model for the assessment of measurements of O3, NO2 and PM.

Previous studies under the European Topic Centre on Air pollution and Climate change Mitigation (ETC/ACM) on the link between trends and meteorology have shown that these links could be estimated by CTMs (chemical transport models). CTMs are useful tools for explaining pollutant trends in terms of the separate impact of individual physio-chemical drivers, such as emissions, and meteorology. It requires, however, multi-year calculations with CTMs designed in specific ways to allow the subtraction of various model scenarios. The method could also be sensitive to the years selected for calculating the perturbations in boundary conditions and meteorology and to uncertainties in emission data. The statistical GAM model that has been developed under ETC/ACM and European Topic Centre on Air pollution, noise, transport and industrial pollution (ETC/ATNI) provides a complementary method for separating the influence of meteorological variability from other processes. This model represents a completely different approach that is based only on observed links between local meteorological parameters (like temperature, wind, etc) and observed pollutant concentration levels. Thus, the model does not contain any representation of the real processes in the atmosphere. We found clear differences in model performance both with respect to geographical area and atmospheric species. In general, the best performance was found for O3 (although not for the peak levels) with gradually lower performance for NO2, PM10 and PM2.5 in that order. With respect to area, the model generally produced the best predictions for Central Europe (Germany, Netherlands, Belgium, France, Austria, Czech Republic) and the poorest for (southern Europe). For wintertime NO2, particularly poor agreement between the GAM model and the measurements was found for the North Italian region. The model agreement for southern Europe and the Iberian Peninsula was also fairly low, although it is variable from site to site. The number of stations with measurements of PM10 and PM2.5 with sufficient length was substantially lower than for O3 and NO2, and thus, a region-by-region comparison of the model performance was not really possible. In general, the PM10 data indicated a better agreement between the model and the measurements for summer than for winter. Furthermore, the GAM model seemed to perform better for the background urban than for rural sites. Poorest performance for PM10 was found for background rural sites in winter. Withdrawing the meteorological factor by the GAM model can help in identifying significant trends. The main reason for this is that meteorology introduces a year-to-year variability which could mask the underlying trend. Meteorology could also induce a trend in the concentrations but this is a matter of length of the timeseries. For short time periods (typically less than 10 years) the variations in meteorology could lead to spurious effects reflecting the weather conditions at the start and end year. On a long timescale the effects of climate change (trends in temperature, precipitation etc) will certainly lead to trends in the concentration of pollutants, but this is beyond the scope of this report. For rural ozone, a statistically significant decline is calculated in the meteorologically adjusted trends in all regions except for the inflow region at the north-western boundary of Europe that shows only minor variations during the 2000-2017 period. Such a decline is also found in the non-adjusted record, but then it is non-significant because of the meteorological variability. For the 2000-2017 period we could, however, not conclude that meteorology alone has caused significant trends in the ozone levels. Some regions show a steady decline in ozone while others show a curve peaking in the early 2000s. Many of the regions indicate a flattening of the ozone trend in the last part of the period. The meteorology adjusted trends for NO2 show a similar pattern as for O3 with decreasing levels in all regions. As for O3 the NO2 trends are seen as a steady decline in some regions and a curve peaking in the early 2000s in other regions. Marked downward meteorology adjusted trends are found for PM10 as well a substantial variability from year to year caused by meteorology. The GAM model estimates a significant meteorologically induced increase in PM10 in the North Italian region, but the number of stations is too few to make strict conclusions on this trend. For PM2.5 the amount of data is insufficient and the monitoring history is considerably shorter and thus we focus on the period 2008-2017 only. A small downward meteorology adjusted trend is estimated for PM2.5, except for the Iberian Peninsula region. To sum up, the GAM model is now in a phase that makes it ready for implementation and use on a regular (annual basis) for areas where its performance is satisfactory.

23 Jun 2020

Prepared by:

Sverre Solberg (NILU), Sam-Erik Walker (NILU), Cristina Guerreiro (NILU) and Augustin Colette (INERIS)

The current report provides a short overview of previous years’ studies on long-term trends in O3, NO2 and PM and the role of meteorological variability for the concentration of these pollutants. The previous studies on the link between trends and meteorology has shown that these links could be estimated by a careful design of model setups using CTMs (chemical transport models). The conclusions from this work is that CTMs are certainly useful tools for explaining pollutant trends in terms of the separate impact of individual physio-chemical drivers such as emissions and meteorology although computationally demanding. The statistical GAM model that have been developed as part of the recent ETC/ACM and ETC/ATNI tasks could be considered as complementary to the use of CTMs for separating the influence of meteorological variability from other processes. The main limitation of the statistical model is that it contains no parameterisation of the real physio-chemical processes and secondly, that it relies on a local assumption, i.e. that the observed daily concentrations could be estimated based on the local meteorological data. We found clear differences in model performance both with respect to geographical area and atmospheric species. In general, the best performance was found for O3 (although not for peak levels) with gradually lower performance for NO2, PM10 and PM2.5 in that order. With respect to area, the model produced the best predictions for Central Europe (Germany, Netherlands, Belgium, France, Austria, Czech Republic) and poorer agreement with observations in southern Europe. Although the GAM model did not detect many  meteorology induced long-term trends in the data, the model is well suited for separating the influence of meteorology from the other driving forces, such as emissions and boundary conditions. The GAM model thus provides robust and smooth long-term trend functions corrected for meteorology as well as the perturbations from year to year, reflecting the variability in weather conditions. One could consider to define a set of performance criteria to decide if the GAM model is applicable for a specific station and parameter.