Uncertainties in emission inventories Wilfried Winiwarter Joint TFEIP & TFMM workshop on uncertainties in emission inventories and atmospheric models Dublin, October 22, 2007
Why consider uncertainties? Uncertainty assessment as a requirement Scientists like it Uncertainty assessment helps identify priorities in further work Performance review of measures taken requires knowledge on method reliability © systems research
Regulatory requirements (Experience from Austria) Uncertainty assessment embedded in QA/QC program Methodological inventory development routinely coupled with uncertainty analysis Inventory improvement (also) based on a-priori uncertainty information: priorities set to assess more uncertain parameters Inventory uncertainty is not used to qualify inventory data (no posterior use) © systems research
Method Understand source of uncertainty: natural variability, unc. of measurement, inapplicability of model Statistical vs. systematic uncertainty (& gross error) Uncertainty sampling Combination of uncertainty Output as a function of one input parameter: Sensitivity analysis Output as a function of all input parameters: Uncertainty analysis © systems research
Uncertainty sampling Measured variations Discrepancy in literature Expert elicitation “reasonable” upper and lower limits, best estimate (equivalent with 95% criterion, removing outliers, will yield µ +/- 2s) Proper distribution may affect resulting distribution, but will influence result only marginally Feedback to QA/QC program © systems research
Uncertainty sheet (NatAir project) parameter Best estimate high low quality Further comments, annotations Emission factor* 24 kg/km² 90 6 E Assuming previous figures to estimate spread (factor 3.75) Activity EU25* 3.82 M km² 4.24 3.40 A Data variability as difference between PBAP area and total area Activity NATAIR domain* 11.79 M km² 14.78 8.80 See above Other parameters Fraction cellulose (debris) 25% 10% 50% Fraction fungal spores 75% 90% Seasonal pattern Totals Total emissions EU 25 92 Gg 350 25 Considering EF uncertainty only Total emissions NATAIR domain 283 Gg 1060 75 © systems research
Sensitivity analysis Assess which parameters contribute to overall uncertainty Important tool to prioritize improvement efforts But: often highly uncertain parameters are simply not accessible Pedigree analysis (van der Sluijs, 2007): independent data quality assessment to understand Discrepancy in literature © systems research
Error propagation … Emission calculation is simple Standard mathematical treatment Monte-Carlo methods (additive terms) (multiplicative terms) s ... standard deviation, RSD ... relative standard deviation s/x © systems research
Error propagation … Error propagation algorithms work as well as Monte-Carlo methods do … … as long as correlation is adequately addressed. Error propagation works for uncorrelated (independent) variables: Note: additive terms allow for overall decrease of relative uncertainty Implicit error reduction: slice a problem into small pieces © systems research
Error propagation for correlated input Transformation required to remove correlated parameters from calculation: E = EF1 * A1 + EF2 * A2 + EF2 * A3 + … E = EF1 * A1 + EF2 ( A2 + A3) + … Note: Uncertainty decrease diminishes (especially if – in the above example – the major uncertainty is with EF) © systems research
Correlated parameters in practice Methane emissions from combustion E = g1 * EF1 * A1 + g1 * EF2 * A2 + g1 * EF3 * A3 + … E = g1 * (EF1 * A1 + EF2 * A2 + EF3 * A3 + …) Note: Despite of apparently different EF’s, the largest share of uncertainty (g1 as fraction of HC measured considered methane) is maintained due to correlation Typical also for VOC species in total HC PM fractions in TSP HM in TSP Possibly also connected with systematic errors © systems research
Reported uncertainty ranges compound Uncertainty range (+/- 2 s; in %) SO2 4 (-10) NOx 12 NMVOC, NH3 20 (-30) CO2 1-2 CH4 15-30 N2O 30-200 Traffic NOx, VOC 30-50 Biogenic VOC +/- factor 4 Sources: Rypdal, 2002; Schöpp et al., 2005; Keizer et al., 2006; Kühlwein&Friedrich, 2000; Leitao et al., 2007 © systems research
Spatial (temporal) assignment and uncertainties Uncertainty due to quality and applicability of surrogate Variations differ by grid cell About two thirds of grid cells display differences not larger than those expected from “plain” uncertainty calculation approx. doubling of uncertainty Geostatistical methods applied allow to identify that differences are spatially correlated surrogate explains only part of spatial variability Sources: Winiwarter et al., 2003; Horabik&Nahorski, 2007 © systems research
Comparison of data sets Validation Performance review 3 1 2 © systems research
Results Uncertainty is small when emission factor is well defined, activity statistics are reliable: CO2 uncertainty associated with GHG emissions is small Uncertainty becomes large when “problem slicing” does not work: PM fractions, HM, POP’s, VOC split, N2O Uncertainty becomes large when underlying processes are not understood well VOC from forests; NO and N2O from soils Spatial (temporal) variability © systems research
Achievements Sectors most strongly contributing to uncertainty Robustness of inventory results: fit-for-purpose? Uncertainty must not compromise inventory consistency (i.e., remain with one “best estimate” result to allow reproduction of the inventory calculations) © systems research
Recommendations Let uncertainty analysis drive your QA/QC program Let sensitivity analysis drive your improvement program Use inventory uncertainty as a reason to focus on key sectors Chapter authors: 8 - Joe Mangino 7 - Kristin Rypdal A2 - Jos Olivier A1- Ian Galbally 6 - Simon Eggleston © systems research